Real-Time Source-Free Object Detection Sairam VCR1 , Varun Gopal1 , Poornima Jain1 , Vineeth N Balasubramanian1,2 , and Muhammad Haris Khan3 IIT Hyderabad, India 2 Microsoft Research 3 MBZUAI {ai20resch13001@,co22btech11015@,ai24resch11002@, vineethnb@cse.}iith.ac.in, [email protected]
arXiv:2606.31834v1 [cs.CV] 30 Jun 2026
1
Abstract. Real-world detectors for autonomous driving, surveillance, and robotics must handle domain-shifts under strict latency and memory constraints, yet existing source-free object detection (SFOD) methods rely on heavyweight architectures that prioritize accuracy alone. We show this trade-off is unnecessary: building on YOLOv10, an NMSfree dual-head detector, we achieve state-of-the-art adaptation accuracy while being faster and more compact. We observe that directly applying vanilla mean-teacher self-training to dual-head detectors leads to suboptimal adaptation performance due to two key factors. First, simple pseudo-label generation strategies, such as using a single head or directly combining high-confidence predictions from both heads, yield suboptimal supervision under domain-shift. We propose DHF (Dual-Head Pseudo-Label Fusion) which selectively admits one-to-one (O2O) and one-to-many (O2M) head predictions, preserving precision and recovering missed objects. Second, we observe domain-shift collapses multi-scale feature discriminability. We propose the use of our MARD (Multi-scale Adaptive Representation Diversification) loss which mitigates this by enforcing detection-aware variance and covariance constraints on multiscale feature maps. Both modules are training-time only, leaving inference unchanged. Across domain-shift benchmarks, our method, RTSFOD yields 1.4 to 3.5% mAP gains, 1.3× higher throughput, with ∼2× fewer parameters than prior state-of-the-art SFOD methods, thus advancing the Pareto frontier of the speed-accuracy-model size trade-off. We report main results with YOLOv10, and demonstrate generalizability with additional YOLO- and DETR-based dual-head detectors. Code is available here: https://github.com/Sairam13001/RT-SFOD/. Keywords: Source-Free Object Detection · Efficient Object Detection
1
Introduction
Source-Free Object Detection (SFOD) adapts a detector trained on labeled source data to an unlabeled target domain without using source samples [43, 12, 42, 51]. This setting is increasingly relevant in applications such as autonomous driving, surveillance, robotics, and medical imaging, where privacy concerns, proprietary restrictions, or storage limitations preclude access to source
2
Sairam VCR et al.
Fig. 1: Accuracy-speed-size trade-off in SFOD. We compare state-of-theart SFOD methods on Cityscapes → Foggy Cityscapes domain-shift, in terms of mAP (%) and inference speed (FPS), with model size (# parameters) encoded by color-graded bubble/star size. Our method achieves the best trade-off across all model-scales, obtaining 53.8% mAP with the large model while remaining the fastest (233 FPS) and most compact (7.2 M params) with the small model. data[36, 33, 29, 1]. Such real-time applications typically operate under strict latency and memory constraints[27, 35]. However, following the pioneering SFOD method [22], existing works[43, 24, 12, 52, 42] largely build on two-stage Faster R-CNN [31], with a few recent ones[19, 52, 2, 51] relying on transformer-based detectors [61, 50]. While these backbones are strong, they are ill-suited for realtime applications [46, 32]. One-stage detectors such as YOLO [39, 15, 44] are designed for efficiency, but limited attention has been given to developing SFOD methods based on them [41]. To bridge this gap, we build a new SFOD framework on YOLOv10[44], which beats the existing methods on accuracy while using a fraction of their parameter count, thus establishing a new state-of-the-art on the accuracy-speed-size frontier as shown in Fig. 1. While we adopt YOLOv10 as the primary architecture for all main experiments, our framework applies to any NMS-free dual-head detector, as we validate on other YOLO- and DETR-based detectors in Table 6. YOLOv10 is the first non-maximum suppression (NMS)-free YOLO detector, employing a dual-head training strategy with one-to-many (O2M) and one-toone (O2O) heads under a consistent matching scheme to enable true end-to-end, real-time detection without post-processing overhead, while maintaining high accuracy. This makes YOLOv10 appear to be an ideal backbone for real-time SFOD; however, unlike Faster R-CNN and transformer-based detectors, it does not explicitly generate region proposals or object queries that provide structured object-level representations for cross-domain alignment, thereby making sourcefree adaptation more challenging. The Mean-Teacher (MT) framework [38] facilitates training without groundtruth labels through teacher-student consistency, and has therefore become a
Real-Time Source-Free Object Detection
(a) Pseudo-label quality
3
(b) Feature discriminability
Fig. 2: Motivation for our proposed modules on the Cityscapes → Foggy Cityscapes benchmark. (a) O2O pseudo-labels are precise but miss many objects; O2M pseudo-labels with standard post-processing provide broader coverage but introduce additional noise; our proposed DHF achieves the best F1-score. (b) Effective rank of YOLOv10 multi-scale PAN features (P3, P4, P5) under domain-shift. The dashed blue line marks the source oracle: effective rank of the source model on Cityscapes val set feature maps. The dotted green line marks the target oracle: effective rank of the target model on Foggy Cityscapes val set feature maps. (No Adaptation) denotes the effective rank of the source model on Foggy Cityscapes val set feature maps. MARD consistently recovers the effective rank best across all scales, closing the gap toward the target oracle. fundamental component of state-of-the-art SFOD methods [12, 52, 42, 51]. However, we observe that a vanilla MT on YOLOv10 is suboptimal for two reasons: (i) As shown in Fig. 2a, the O2O head produces high-precision (0.923) pseudolabels but tends toward low recall (0.446), missing objects. In contrast, use of the O2M head for pseudo-labeling improves coverage (recall 0.469) but introduces additional noise (precision 0.847). Using either head alone for pseudo-labeling, or directly combining head predictions remains suboptimal (Tab. 5b). (ii) The dualhead consistent assignment mechanism benefits from highly discriminative features [44], however, we hypothesize that domain-shift erodes the discriminability. To verify this, for each scale level ℓ, we reshape the feature map Fℓ ∈ RCℓ ×Hℓ ×Wℓ into a spatially flattened matrix Zℓ ∈ R(Hℓ Wℓ )×Cℓ , where rows correspond to spatial locations and columns correspond to P channels. We compute singular values {σi } of Zℓ , normalize them as p = σ / i i j σj , and define effective rank as P erankℓ = exp(− i pi log pi ), which quantifies the effective number of channel dimensions carrying independent information. Over a dataset of N images, we PN (n) calculate the dataset-level mean effective rank as erankℓ = N1 n=1 erankℓ , and ℓ report its normalized percentage as erank%ℓ = 100 × erank in Fig.2b. DomainCℓ shift erodes the effective rank of features across scales as observed in Fig.2b (No Adaptation) , indicating the collapse of features toward lower-dimensional subspaces. Standard mean-teacher self-training (Adapted w/o MARD) recovers less than half of this lost diversity, placing a ceiling on adaptation performance. To address these two bottlenecks, we introduce two modules respectively: (DHF) Dual-Head Pseudo-Label Fusion. Treating duplicate-free highprecision O2O predictions as anchors, we selectively supplement them with non-
4
Sairam VCR et al.
redundant, high-scoring O2M predictions to recover objects that O2O misses. Fig.2a shows how our DHF loss increases F-1 score by 10.7%, leading to a gain in adaptation performance (Tab.5b). This fusion is applied only during trainingtime pseudo-label construction; at inference, the detector retains its original NMS-free O2O-only inference with no additional overhead. (MARD) Multiscale Adaptive Representation Diversification. MARD enforces structured representational constraints on the multi-scale features, preserving the discriminative capacity (Fig.2b (Adapted w/ MARD) ) that domain-shift otherwise erodes. Together, these components facilitate stable adaptation of YOLOv10, achieving similar or higher accuracy than state-of-the-art approaches, with 1.3x higher FPS and requiring significantly fewer parameters. Our main contributions are: – We present, to the best of our knowledge, the first systematic study of NMSfree dual-head detectors for SFOD. Our proposed method, RT-SFOD, establishes a new state-of-the-art on the accuracy-speed-size frontier. – We propose DHF, a training-time fusion of high-precision O2O and selectively filtered O2M predictions for superior pseudo-label quality, and MARD, a multi-scale feature diversification mechanism that counteracts feature rank degradation under domain-shift. – Across domain-shift scenarios, our proposed modules yield state-of-the-art or competitive adaptation accuracy, while the efficient YOLOv10 backbone contributes 1.3x higher throughput with ∼2x fewer parameters than prior SFOD methods.
2
Related Works
Source-Free Object Detection: SFOD was introduced in [22], which identified noisy pseudo-labels as the main challenge and addressed it using entropybased confidence thresholding and false-negative augmentation. LODS [21] applied bidirectional knowledge distillation on original and style-transferred target images with graph-based feature alignment to learn domain-invariant representations. [8] proposed A2 SFOD, which uses variance-based target splitting and adversarial alignment in a mean-teacher framework. IRG [43] adopted a mean-teacher framework with graph CNN-guided contrastive learning for target feature alignment. Simple-SFOD [12] adapted batch normalization statistics and leveraged strong augmentations with fixed pseudo-labels in a teacherstudent setup for domain adaptation. FALCON-SFOD [42] showed that domainshift weakens object-focus in features and leveraged foundation model priors to address this issue. Following [22], all above methods adopt Faster R-CNN as the base detector. Recent SFOD [19, 52] works utilize Detection Transformers (DETRs). In [19], a historical student loss is used to stabilize student-training and mitigate noisy pseudo-labels. FRANCK [52] is an SFOD framework that enhances DETR adaptation via category-level contrastive learning, instance reweighting, and uncertainty-weighted feature distillation. Faster R-CNN and DETR are computationally intensive [28], and SFOD methods built on them focus mainly on adaptation accuracy, overlooking real-time performance metrics such as FPS and model size.
Real-Time Source-Free Object Detection
5
Unsupervised Domain Adaptive Object Detection: UDAOD addresses domain-shift using both labeled source and unlabeled target data. Early methods focused on Faster R-CNN [4, 7, 10, 49], while later works adopted DETRbased models: MTM [48] proposed two-stage DETR adaptation with source pretraining on style-transferred images and masked feature alignment; MTTrans [55] applied a ResNet50-based DETR in a mean-teacher framework; DA-DETR [57] and SFA [47] introduced transformer detectors with dedicated alignment modules; and BiADT [25] leveraged box-coordinate queries as positional priors using a DAB-DETR backbone. UDOAD requires access to labeled source data, which limits its applicability in real-world applications where source data privacy and unavailability are major concerns, shifting the focus more towards SFOD. Efficiency-centric Object Detection: Research on fast/lightweight detectors includes SSD-MobileNet [26, 13] and EfficientDet [37]. NanoDet [30] adopts depthwise convolutions with an anchor-free design, while the YOLO line [39, 45, 44, 18] has been progressively optimized for speed. Real-time transformer detectors include Real-Time-DETR [59], RF-DETR [32], LightWeightDETR [6], and Quantized RT-DETR [14], but they generally trail YOLO in throughput. For instance, [44] reports RT-DETR-R18 at ∼46.5 AP on COCO with ∼4.6 ms/frame (∼217 FPS), whereas YOLOv10-S achieves similar AP (46.3) in 2.49 ms/frame (∼401 FPS), about 1.8× faster. The same study [44] also shows Faster R-CNN and DETR variants have higher compute/latency, while YOLOv10’s streamlined heads and lightweight blocks surpass YOLOv8 at higher accuracy with ∼2× fewer parameters, making it especially efficient for real-time systems. SF-YOLO [41] is the first work to apply real-time detectors to SFOD. It adapts a single-head, NMS-based YOLO detector (YOLOv5 [39]) using meanteacher self-training with student stabilization and learned augmentations. In contrast, our work focuses on NMS-free dual-head detectors, which bring new challenges beyond those considered in SF-YOLO, such as pseudo-label precisionrecall trade-off across heads. We further provide the evidence of a previously unreported phenomenon: feature discriminability collapse under domain-shift; and provide a scale-aware regularization loss to mitigate this effect and preserve feature diversity. To our knowledge, no prior work systematically studies efficiencydriven SFOD for NMS-free detectors, leaving newer efficient architectures such as YOLOv10 [44] underexplored. We fill this gap with an SFOD framework for NMS-free dual-head detectors, instantiated primarily on YOLOv10, where our proposed modules drive adaptation accuracy gains while the throughput and latency advantages stem from the efficient YOLO backbone. We also demonstrate generality on other YOLO- and DETR-based dual-head detectors.
3
Methodology
3.1
Preliminaries
Problem statement. Let the unlabeled target dataset be Dt = {xi }N i=1 . Given a detector pre-trained on a labeled source domain, our objective is to adapt
6
Sairam VCR et al.
Fig. 3: Overview of RT-SFOD. (i) Mean-Teacher framework with DHF and MARD. (ii) DHF: High-precision O2O predictions serve as anchors; nonredundant O2M boxes with low overlap w.r.t. O2O anchors are selectively added to form the final pseudo-labels. (iii) MARD: (iii-a) Foreground and background feature vectors are sampled from multi-scale features using pseudo boxes and regularized via variance and covariance constraints, (iii-b) increasing perchannel variance for better discriminability and reducing off-diagonal covariance for greater channel diversity. it to Dt without accessing any source samples, while maintaining a favorable accuracy-efficiency trade-off suitable for real-time use. Mean-Teacher self-training. Most state-of-the-art SFOD methods follow the Mean-Teacher (MT) teacher-student paradigm. A teacher detector fθT produces pseudo-labels on the target domain, which are then used to supervise a student detector fθS . For a target image x, we construct a weakly augmented view xw = aw (x) and a strongly augmented view xs = as (x). The teacher predicts pseudolabels on xw , and the student is trained to match them on xs . After each step, the teacher parameters are updated via an exponential moving average (EMA) of the student with momentum µ: θT ← µ θT + (1 − µ) θS . YOLOv10 preliminaries. We adopt YOLOv10 as the primary base detector due to its strong accuracy-efficiency trade-off and end-to-end NMS-free inference, which are desirable for real-time SFOD. YOLOv10 employs two prediction heads trained under a consistent matching scheme: a one-to-one (O2O) head that assigns each object to a single prediction, and a one-to-many (O2M) head that allows multiple matched predictions per object to strengthen supervision. YOLOv10 operates on a multi-scale feature pyramid (PAN); we denote the stu-
Real-Time Source-Free Object Detection
7
Algorithm 1 Proposed Real-time SFOD. Require: AdaBN-initialized checkpoint θ0 , unlabeled target data Dt and hyperparameters 1: Initialize student θS ← θ0 , teacher θT ← θ0 2: for each epoch do 3: for each minibatch x ∼ Dt do 4: Form xw = aw (x) and xs = as (x) 5: Teacher predicts P o and P m on xw 6: Fuse with DHF to obtain pseudo-labels Ŷ w (Eq. 4) 7: Map Ŷ w → Ŷ s via the strong-view geometry 8: Student predicts Y on xs ; compute detection loss Ldet (Ŷ s , Y ) 9: Compute MARD loss Lmard on multi-scale PAN features (Eq. 8) 10: Update θS using Ldet + λLmard 11: end for 12: Update teacher: θT ← µθT + (1 − µ)θS (epoch-level EMA) 13: end for
dent features at pyramid level l as FℓS , for ℓ ∈ {P 3, P 4, P 5}. During inference, only the O2O head is used to produce duplicate-free detections. Why vanilla MT is suboptimal for YOLOv10. Despite being a strong SFOD baseline, vanilla MT underperforms on YOLOv10 for two reasons. (i) Its dual heads yield complementary but mismatched pseudo-labels: O2O is duplicate-free and precise yet misses many objects under domain-shift, whereas use of the O2M head for pseudo-labeling improves coverage but introduces additional noise; using either alone, or directly combining them, results in suboptimal supervision. (ii) Domain-shift degrades YOLOv10’s multi-scale feature discriminability[44], reducing effective rank across pyramid levels (Fig. 2b), and MT self-training only partially recovers this diversity. We therefore propose DHF for higher-quality pseudo-labels and MARD to preserve discriminative features. 3.2
Proposed Real-Time SFOD
Following [12], we warm-start the teacher using Adaptive Batch Normalization (AdaBN): we update BN statistics on target images to obtain a stronger targetinitialized checkpoint. Teacher and student are initialized with this checkpoint before the adaptation begins. After applying AdaBN, we found that updating the teacher model once per epoch yields better performance than updating it at every training step. Algorithm 1 summarizes the adaptation loop. Below, we explain the two proposed modules of our framework. DHF: Dual-Head Pseudo-Label Fusion Let P o and P m denote the teacher predictions from the O2O and O2M heads on the weak view xw , respectively. Each prediction is a tuple p = (b, s, c) where b ∈ R4 denotes box coordinates (x1 , y1 , x2 , y2 ), s is the confidence, and c is the class. We first keep high-confidence O2O predictions as anchors: \hat {Y}^{o} = \{p\in \mathcal {P}^{o} \mid s(p)\ge \tau _o\}.
(1)
8
Sairam VCR et al.
We then consider O2M candidates C m = {p ∈ P m | s(p) ≥ τm }, and add only those that have low overlap w.r.t. the O2O anchors: \mathcal {E} = \left \{p\in \mathcal {C}^{m} \;\middle |\; \max _{q\in \hat {Y}^{o}} \mathrm {IoU}(b_p,b_q) \le \tau _{\text {no}} \right \}. \label {eq:extras}
(2)
This rule uses the O2O head as a high-precision prior and lets O2M contribute only non-redundant boxes, improving pseudo-label coverage while mitigating noisy supervision during self-training. Because O2M predictions can contain duplicates among themselves, we apply class-wise NMS only on the O2M extras: \mathcal {E} \leftarrow \mathrm {NMS}(\mathcal {E},\tau _{\text {dup}}).
(3)
Finally, the fused pseudo-label set is \label {eq:DHF} \hat {Y}^w = \hat {Y}^{o} \cup \mathcal {E}.
(4)
MARD: Multi-scale Adaptive Representation Diversification Let {FℓS }ℓ∈{3,4,5} be the three feature maps input to the YOLOv10 detection head (PAN outputs), where Fℓ ∈ RB×Cℓ ×Hℓ ×Wℓ . We compute MARD on these features during the student forward pass on xs . Crucially, MARD gradients propagate through the PAN into the backbone, actively shaping feature learning during adaptation. Unlike representation-learning methods that regularize global embeddings[3], MARD constructs its feature vector set using detection structure: – Pseudo-box guided foreground feature vectors. From the pseudolabels Ŷ w , we sample K spatial locations inside each pseudo box and extract the corresponding vectors from the detection feature map Fℓ . This ties MARD to object evidence used by the detector. – Complementary background feature vectors. We also sample M locations from background regions (valid, non-padded pixels outside pseudo boxes), encouraging object-background separability in the same feature space. – Noise control under SFOD. We restrict sampling to the top-Kb pseudo boxes with confidence above τb . This improves robustness to pseudo-label noise, and the adaptive weighting in Eq. (9) further downweights MARD when pseudo-label quality is low. YOLO-style detectors distribute predictions across feature pyramid levels (PAN) according to object size. To keep MARD detection consistent, we assign each p pseudo box b to a feature level based on its size s(b) = w(b)h(b) in pixels and stride-derived thresholds:
g(b)= \begin {cases} 3 & s(b)\le \eta \cdot \mathrm {stride}_3,\\ 4 & \eta \cdot \mathrm {stride}_3 < s(b)\le \eta \cdot \mathrm {stride}_4,\\ 5 & s(b) > \eta \cdot \mathrm {stride}_4, \end {cases} \label {eq:levelassign}
(5)
where η is a constant and strideℓ is the effective stride of level ℓ. We then sample foreground/background feature vectors per level using only boxes assigned to
Real-Time Source-Free Object Detection
9
that level. This prevents MARD from mixing incompatible scales and directly targets the scale-specific features used for small/medium/large objects. For each level ℓ, we collect a feature vector matrix Zℓ ∈ RNℓ ×Cℓ , containing foreground and background feature vectors. We then apply two complementary constraints: (i) a variance term to prevent feature collapse (ensuring each channel maintains a non-trivial spread), and (ii) a covariance term to reduce redundancy across channels. \mathcal {L}_{\text {var}}(Z_\ell ) &= \frac {1}{C}\sum _{c=1}^{C} \max \left (0,\gamma -\sqrt {\mathrm {Var}(Z_{\ell ,:,c})+\epsilon }\right ), \label {eq:var} \\ \mathcal {L}_{\text {cov}}(Z_\ell ) &= \frac {1}{C(C-1)}\sum _{i\neq j} \left (\mathrm {Cov}(\tilde {Z}_\ell )_{ij}\right )^2, \label {eq:cov} (7) where Z̃ℓ denotes channel-wise normalized tokens, γ is a hyperparameter and ϵ is a small constant for stability. While the variance and covariance terms share a functional form with collapse-prevention objectives used in self-supervised learning [3], MARD differs in that it operates on multi-scale PAN features rather than global embeddings, constructs feature vector sets via pseudo-box guided foreground and background sampling tied to detection structure, and enforces scale-consistent regularization via the level assignment in Eq. (5). The MARD loss is defined as: \mathcal {L}_{\text {mard}} = \sum _{\ell \in \{3,4,5\}} \left ( \alpha \,\mathcal {L}_{\text {var}}(Z_\ell ) + \beta \,\mathcal {L}_{\text {cov}}(Z_\ell ) \right ). \label {eq:mard}
(8)
To avoid over-regularizing when pseudo-labels are unreliable, we use an adaptive weight \lambda (t) = \lambda _0 \cdot \mathrm {ramp}(t) \cdot \mathrm {gate}(\bar {s}), \label {eq:lambda} (9) where ramp(t) linearly increases from 0 to 1 over a warmup period, and gate(s̄) increases with the batch-average pseudo-label confidence s̄ (clipped to [0, 1]). We also compute MARD periodically (every I steps) to limit overhead. 3.3
Training Objective and Inference
Given the student outputs on xs and pseudo-labels Ŷ w mapped to strong-view Ŷ s , we first calculate the standard YOLOv10 detection loss (box regression, classification, and distribution focal loss), computed over both the O2O and O2M student heads to retain the dual-head training dynamics of YOLOv10: \mathcal {L}_{\text {det}} = \mathcal {L}_{\text {box}} + \mathcal {L}_{\text {cls}} + \mathcal {L}_{\text {dfl}}.
(10)
Total loss. Our final optimization objective is: \mathcal {L} = \mathcal {L}_{\text {det}} + \lambda (t)\,\mathcal {L}_{\text {mard}}. \label {eq:total}
(11)
Inference-time efficiency. Dual-Head Fusion and MARD are used only during training-time adaptation. After adaptation, we deploy the student fθS as a standard YOLOv10 detector, preserving end-to-end NMS-free inference and incurring no additional runtime cost.
10
Sairam VCR et al.
Table 1: Comparison with state-of-the-art on C2F (Cityscapes → Foggy Cityscapes). Bold denotes the best SFOD result; underline the second-best. Efficiency metrics are measured using each method’s official public codebase. † FRANCK has no public implementation; its efficiency figures are adopted from DRU, which shares the same base detector. S = Source-only; UDAOD = Unsupervised DA OD; SFOD = Source-Free OD. All our results are averaged over three random seeds. Category Method
Base
S
YOLOv10S YOLOv10M YOLOv10L
7.2 15.4 24.4
233 105 67
4.3 9.5 15.0
41.0 46.1 49.6 12.0 27.0 4.9 43.6 47.8 50.0 12.8 29.8 5.0 47.1 52.4 58.5 18.6 35.1 10.8
17.1 20.1 24.5
39.0 40.2 44.8
29.6 31.2 36.5
FRCNN DETR FRCNN DETR FRCNN
NA NA NA NA NA
NA NA NA NA NA
NA NA NA NA NA
43.7 52.8 44.6 61.6 58.5
49.3 54.5 49.7 35.4 54.8
35.2 41.0 44.9 43.6 47.1
47.9 47.1 53.0 55.9 59.9
47.4 51.2 52.5 53.4 57.5
34.0 44.2 43.8 43.8 7.2 46.5 40.0 40.0† 41.0 7.2 15.4 24.4
51 48 42 42 172 52 29 29† 27 233 105 67
19.5 20.6 23.8 23.8 5.8 19.1 35.0 35† 37.0 4.3 9.5 15
37.4 45.2 51.9 24.4 39.6 25.2 42.0 48.7 56.3 19.3 39.3 5.5 40.9 48.0 58.9 29.6 51.9 50.2 41.0 48.3 58.7 33.6 54.8 54.3 48.6 51.8 64.5 24.7 48.1 37.6 55.5 58.0 71.5 36.6 53.7 46.1 48.3 51.5 62.5 26.2 43.2 34.1 48.1 49.3 60.6 33.9 48.2 36.9 48.8 51.0 62.8 36.6 52.7 39.9 47.6 52.6 64.8 25.9 46.0 41.5 53.7 57.4 67.3 26.1 46.3 44.9 58.4 64.7 71.0 38.5 57.0 49.5
31.5 34.2 36.2 38.6 23.1 40.6 34.2 34.0 36.4 30.4 35.8 39.0
41.6 41.6 44.1 46.2 44.8 50.5 48.6 47.9 48.2 45.2 47.3 52.1
37.1 35.9 45.0 46.9 42.9 51.6 43.6 44.9 47.1 44.3 47.4 53.8
Source only
AT (CVPR’22) MRT (ICCV’23) UDAOD CAT (CVPR’24) DATR (TIP’25) SEEN-DA (CVPR’25)
SFOD
IRG (CVPR’23) PETS (ICCV’23) FRCNN Simple-SFOD (ECCV’24) FALCON-SFOD (CVPR’26) SF-YOLO (ECCV’24 ws) YOLOv5S SF-YOLO (ECCV’24 ws) YOLOv5L DRU (ECCV’24) Def-DETR FRANCK (TIP’25) VFM-SFOD (AAAI’26) RT-SFOD (Ours)
4
YOLOv10S YOLOv10M YOLOv10L
Params (M) FPS Lat. (ms) prsn rider car truck bus train mcycle bicycle mAP
54.1 51.7 57.1 60.4 64.5
62.3 68.7 63.7 74.3 71.7
31.9 35.9 40.8 35.7 42.0
54.4 58.1 66.0 60.3 61.2
Experiments
Datasets and Metrics. Following existing works, we use five publicly available datasets covering four domain-shift scenarios: Cityscapes[9], Foggy Cityscapes[34], KITTI[11], Sim10k[16], and BDD100k[54]. Detailed description of datasets is provided in supplementary (Sec. S.4.2). Unlike existing SFOD methods, which only report target domain accuracy (mean average precision mAP at 0.5 IoU threshold), we report mAP along with efficiency metics like throughout (FPS), number of parameters, and latency (ms/img). As noted in [32], many efficiency-focused methods [59, 41, 44] report FP32 PyTorch accuracy but FP16 TensorRT efficiency, leading to an unfair comparison. Moreover, [32] shows that naive FP16 quantization can significantly degrade performance. Therefore, we report all primary results in FP32 using PyTorch, and provide FP16 TensorRT results in the supplementary (Sec. S.8) for comparability. Implementation details. The student is trained using SGD with an initial learning rate of 1e-4 and cosine annealing, gradient clipping at norm 10, and batch size 16 for 60 epochs. The teacher is updated via Exponential Moving Average (EMA) with momentum 0.999 at the end of each epoch. All experiments are conducted on a single NVIDIA RTX A6000 GPU. We use the same hyperparameters across all domain-shifts. Details about hyperparameters are included in the supplementary (Sec. S.4.1).
Real-Time Source-Free Object Detection
4.1
11
Comparison with State-of-the-Art Methods
Following existing works, we compare our method with (i)UDA-based object detection (UDAOD), which retains source data during adaptation [23, 60, 17, 5, 20], and (ii) SFOD methods [43, 24, 12, 19, 52, 42, 51, 41]. Cityscapes → Foggy Cityscapes (C2F). Table 2: S2C and K2C compariTable 1 presents results on the most widely son. Efficiency metrics are identiused SFOD benchmark. Our RT-SFOD-L cal to those in Table 1 achieves 53.8 mAP, surpassing the previCat. Method Base S2C K2C ous best SF-YOLO-L (51.6) by +2.2 mAP YOLOv10S 51.2 37.4 and remaining competitive with UDAOD S Source only YOLOv10M 53.5 40.3 YOLOv10L 56.5 45.0 methods (MRT: 51.2, DATR: 53.4) unDATR[5] DETR 64.3 53.3 UDA der the stricter source-free constraint. Our SEEN-DA[20] FRCNN 66.8 67.1 RT-SFOD-M (47.4 mAP) outperforms most IRG[43] 45.2 46.9 PETS[24] 57.8 47.0 Faster R-CNN and Deformable DETR based FRCNN LPLD[53] 49.4 51.3 Simple[12] 55.4 46.2 SFOD approaches. Compared to SF-YOLOFALCON[42] 58.8 50.1 SFOD SF-YOLO[41] YOLOv5S 57.7 50.6 L, RT-SFOD-L is 1.3× faster (67 vs. 52 SF-YOLO[41] YOLOv5L 69.8 63.7 DRU[19] 58.7 45.1 FPS) and 1.9× smaller (24.4M vs. 46.5M DefDETR 63.1 48.5 FRANCK[52] VFM[51] 67.4 54.7 parameters). Against Deformable DETRYOLOv10S 59.8 52.3 based methods (DRU, FRANCK, VFMRT-SFOD (Ours) YOLOv10M 66.4 54.2 YOLOv10L 71.2 60.9 SFOD), which cluster around 27–29 FPS and 40–41M parameters, RT-SFOD-M matches their best accuracy at 3.6× higher throughput and 2.6× fewer parameters. Our smallest variant, RTSFOD-S (44.3 mAP), is on par with DRU (43.6) and FRANCK (44.9) while running at 233 FPS with only 7.2M parameters. These trade-offs are visualized in Fig. 1. Sim10k→Cityscapes (S2C) and KITTI →Cityscapes (K2C). Following prior works, we report AP(on Car) for S2C and K2C. On S2C, our RT-SFODL achieves 71.2 mAP, outperforming SF-YOLO-L (69.8) by +1.4 and VFMSFOD (67.4) by +3.8. On K2C, it reaches 60.9 mAP, second only to SF-YOLO-L (63.7), while using 1.9× fewer parameters and 1.3× higher throughput. We attribute the gap to SF-YOLO-L to model capacity: SF-YOLO-L uses 1.9× more parameters (46.5M vs. 24.4M), and at matched scale (both 7.2M), RTSFOD-S outperforms SF-YOLO-S (52.3 vs. 50.6). Moreover, RT-SFOD-S/M outperform all five Faster R-CNN methods and two of three DETR-based methods on K2C while being faster and smaller, indicating that this camera-shift scenario particularly benefits from additional model capacity. The S→M→L scaling yields 59.8→66.4→71.2 on S2C and 52.3→54.2→60.9 on K2C, demonstrating consistent gains as model size increases. Cityscapes → BDD100k (C2B). Table 3 presents a challenging large-scale benchmark spanning diverse cities, weather conditions, and times of day. Our RT-SFOD-L achieves 46.5 mAP, exceeding VFM-SFOD (43.0) by +3.5 and FALCON-SFOD (36.9) by +9.6, and also beating both UDAOD baselines. Even RT-SFOD-S (41.1 mAP) is competitive with heavyweight Def-DETR methods such as FRANCK (40.7). The consistent gains across C2F (weather), C2B (large-
12
Sairam VCR et al.
Table 3: Comparison with state-of-the-art on C2B (Cityscapes → BDD100k). Efficiency metrics are identical to those in Table 1. Category Method
Base
truck car rider person motor bicycle bus mAP
S
Source only
YOLOv10S 17.5 59.6 29.1 YOLOv10M 18.8 61.8 32 YOLOv10L 23.1 68.3 35.4
42.4 44.5 49.2
14.6 15.8 17.8
19.3 22.6 24
20 28.9 21.5 31 25.3 34.7
UDA
MRT[60] (ICCV’23) DATR[5] (TIP’25)
Def-DETR Def-DETR
24.7 63.7 30.9 26.9 73.4 42.8
48.4 58.5
20.2 24.2
22.6 37.3
25.5 33.7 39.9 43.3
SFOD
IRG[43] (CVPR’23) 31.4 59.7 32.8 PETS[24] (ICCV’23) 19.3 62.4 34.5 FRCNN Simple-SFOD[12] (ECCV’24) 32.0 60.0 33.4 FALCON-SFOD[42] 32.6 59.8 34.0 DRU[19] (ECCV’24) 27.1 62.7 36.9 Def-DETR 29.5 65.6 43.3 FRANCK[52] VFM-SFOD[51] 33.2 72.3 44.2 YOLOv10S 28.8 71.6 44.8 RT-SFOD (Ours) YOLOv10M 33.7 75.9 47.3 YOLOv10L 36.4 77.0 49.3
39.9 42.6 40.2 40.0 45.8 55.0 54.9 53.8 57.5 58
16.7 17.0 19.7 25.7 22.7 30.0 32.9 28.7 33 34.1
26.9 26.3 29.9 35.7 32.5 28.0 29.0 26.5 28.6 31.4
21.5 32.7 16.9 31.3 24.9 34.3 30.5 36.9 28.1 36.6 33.6 40.7 34.8 43.0 33.6 41.1 35.8 44.5 39.6 46.5
scale diversity), S2C (rendering gap), and K2C (camera shift) indicate that our modules generalize across domain gap characteristics. 4.2
Ablation Study
Table 4: Ablation study of the proposed components with RT-SFOD-M across domain-shifts C2F, S2C, K2C. Module
Components
C2F
S2C K2C
DHF MARD prsn rider car truck bus train mcycle bicycle mAP AP AP Source-trained AdaBN MT (O2M) MT (O2O) MT (O2O) + DHF MT (O2O) + MARD MT (O2O) + DHF + MARD
× × × × ✓ × ✓
× × × × × ✓ ✓
45.6 50.9 50.3 51.5 52.9 51.8 53.7
48.8 50 51.4 57.9 54.5 63.7 55.0 63.8 56.4 67.0 55.7 66.2 57.4 67.3
7.48 15.7 24.3 24.9 25.5 25.2 26.1
27 3 35.8 15.5 45.0 41.1 45.2 41.7 45.6 44.3 45.9 44.7 46.3 44.9
20.1 26.6 33.5 33.8 35.2 35.0 35.8
42.2 45.1 45.3 45.9 46.6 46.2 47.3
30.5 52.5 40.3 37.4 54.2 45.2 44.7 61.4 49.9 45.2 62.1 50.6 46.7 65.1 53.0 46.3 65.3 52.5 47.4 66.4 54.2
Table 4 presents a component-wise ablation with RT-SFOD-M across three benchmarks. Starting from the source-trained model (30.5/52.5/40.3 mAP on C2F/ S2C/K2C), AdaBN provides a better starting point for the teacher by aligning batch normalization statistics before training. Adding mean-teacher (MT) self-training yields further gains: MT with one-to-many (O2M) pseudolabels achieves 44.7 mAP on C2F, while MT with one-to-one (O2O) obtains 45.2. The O2O head’s slight advantage is consistent with our precision-recall analysis in Fig. 2a: while O2M with standard post-processing provides broader coverage (recall 0.469 vs. 0.446), it introduces additional noise (precision 0.847 vs. 0.923) that can compound during iterative self-training. However, despite these gains, the vanilla MT framework remains sub-optimal.
Real-Time Source-Free Object Detection
13
Effect of DHF. Adding Dual-Head Fusion to MT (O2O) yields consistent gains: +1.5 on C2F (45.2 → 46.7), +3.0 on S2C (62.1 → 65.1), and +2.4 on K2C (50.6 → 53.0), with the largest per-class gains on car (+3.2) and train (+2.6). Effect of MARD. Adding MARD loss (w/o DHF) yields +1.1 on C2F (45.2 → 46.3), +3.2 on S2C (62.1 → 65.3), and +1.9 on K2C (50.6 → 52.5). As shown in Fig. 2b, MARD recovers 76–84% of the effective rank lost to domainshift at each PAN-feature level (vs. only 26–48% without it), restoring feature discriminability. Complementarity of DHF and MARD. Combining both modules achieves the best results: 47.4/66.4/54.2 on C2F/S2C/K2C, with total gains of +2.2/ +4.3/+3.6 over MT (O2O). The combined improvement exceeds either module in isolation, confirming they address complementary bottlenecks: DHF improves the supervisory signal while MARD improves learned representations. Table 5: Ablation studies on RT-SFOD-M. Left: MARD sub-objectives. Right: Pseudo-label strategy comparison. (a) Ablation of MARD components. Baseline: MT (O2O) + DHF.
(b) Pseudo-label strategy comparison (RT-SFOD-M, C2F).
Variance Covariance C2F S2C K2C
Strategy
× ✓ × ✓
4.3
× × ✓ ✓
46.7 65.1 53.0 47.2 65.9 53.8 47.0 65.7 53.5 47.4 66.4 54.2
Prec Recall F1 mAP
O2O only 0.923 0.446 0.602 45.2 O2M+NMS 0.847 0.469 0.603 45.2 O2O ∪ O2M+NMS 0.834 0.512 0.634 45.5 DHF (Ours) 0.919 0.525 0.668 46.7
Analysis
MARD sub-component analysis. Table 5a ablates the two MARD components using MT (O2O) + DHF as the fixed baseline. The variance term (preventing channel collapse) provides the larger individual contribution (+0.5/+0.8/+0.8 on C2F/S2C/K2C), consistent with channel collapse being the primary mode of rank degradation in Fig. 2b. The covariance term (reducing inter-channel redundancy) provides a complementary gain (+0.3/+0.6/+0.5). Combining both yields the full MARD improvement (+0.7/+1.3/+1.2), confirming they address distinct failure modes. Pseudo-label fusion strategy comparison. Table 5b compares four pseudolabel strategies on label-quality metrics and downstream mAP. O2O yields high precision (0.923) but low recall (0.446), while O2M + NMS provides broader coverage (0.469) at lower precision (0.847), giving a comparable F1 (0.603). Directly merging O2O with O2M + NMS improves recall (0.512) but reduces precision (0.834), yielding a suboptimal F1 (0.634). Our DHF uses O2O boxes as high-confidence anchors and selectively adds non-redundant, high-scoring O2M boxes with low overlap w.r.t. O2O anchors, retaining near-O2O precision (0.919) while boosting recall to 0.525, achieving the best F1 (0.668) and mAP (46.7). This confirms that pseudo-label selection must balance precision and recall in
14
Sairam VCR et al.
self-training: the direct union improves recall but reduces mAP, as noisy pseudolabels can compound across iterations. Hyperparameter sensitivity. Fig. 4 shows C2F mAP of RT-SFOD-M as a function of four key hyperparameters, with performance stable across all tested ranges (≤1.5 mAP variation). Both confidence thresholds (τo2o , τo2m ) peak at 0.5; the τo2m curve is notably flatter (46.0–47.4), indicating DHF’s IoU-based deduplication provides robustness against the O2M threshold choice. The optimal MARD weight is λ0 =0.05, with Fig. 4: Hyperparameter sensitivity on performance degrading gracefully C2F (RT-SFOD-M). mAP (%) vs. O2O at both extremes; the effective threshold τo2o , O2M threshold τo2m , weight is further modulated by MARD weight λ0 , and EMA momentum µ. a linear warmup and confidence Variation is ≤1.5 mAP in all cases. gate (Sec. 3.2). EMA momentum µ=0.999 balances teacher stability and responsiveness. No hyperparameter requires per-benchmark tuning, and the same values transfer across detectors and scales, likely because they depend on relative statistics (calibrated confidence, geometric IoU, and feature distributions) rather than architecture-specific factors. The full set of values used is: τo2o =τo2m =0.5, τno =0.2, τdup =0.7, λ0 =0.05, K=8, M =128, µ=0.999; full sensitivity analysis is in the supplementary (Sec. S.4.1). Feature Discriminability Analysis. Fig. 5 illustrates the average cosine similarity between each anchor point’s extracted features and all others across PAN scales (P3, P4, and P5) on the Foggy Cityscapes validation set. We compare the Source model, Vanilla Mean-Teacher, and the proposed MARD approach. Notably, MARD consistently yields lower cosine similarity across all scales, indicating more structured feature representations and enhanced discriminability. Fig. 5: The average cosine similarity of each Generality across Dual-Head Deanchor point’s extracted features with all tectors. While our main experiments others across PAN scales(P3, P4, P5) on use YOLOv10, RT-SFOD can be ap- Foggy Cityscapes validation set. plied to any NMS-free dual-head de-
Real-Time Source-Free Object Detection
15
tector. DHF requires only a precise O2O head and a dense, high-recall O2M head, and MARD applies to any multi-scale feature pyramid. Table 6 presents the analysis of RT-SFOD applied with the same hyperparameters to three additional dual-head detectors on C2F. As shown in Table 6, RTTable 6: Generality of RT-SFOD SFOD improves MT (O2O) conacross dual-head detectors on C2F sistently across YOLOv26S/M/L [40] domain-shift. (+2.5/+2.8/+2.1), MS-DETR [58] (+2.0), and Mr. DETR [56] (+2.3), Detector MT (O2O) RT-SFOD matching the gains observed on YOLOv26S [40] 42.1 44.6 (+2.5) YOLOv10. This confirms that the YOLOv26M [40] 46.0 48.8 (+2.8) 50.6 52.7 (+2.1) benefits of DHF and MARD are YOLOv26L [40] 42.3 44.3 (+2.0) properties of the dual-head paradigm MS-DETR [58] Mr. DETR [56] 43.8 46.1 (+2.3) rather than of a specific architecture. For single-head detectors, DHF has no second candidate set and reduces to confidence-threshold pseudo-labeling as in SF-YOLO [41], while MARD remains applicable whenever pseudo-labels and multi-scale features are available.
5
Conclusion
We present RT-SFOD, a source-free object detection framework for NMS-free dual-head detectors, instantiated primarily on YOLOv10, that improves the accuracy-speed-model size trade-off for domain-adaptive detection. We identify two challenges in directly applying mean-teacher self-training to NMSfree dual-head detectors: (i) suboptimal pseudo-label supervision from unconditioned use of O2O and O2M predictions, and (ii) feature rank degradation under domain-shift. To address these, we propose Dual-Head Fusion, which uses high-precision O2O predictions as anchors and selectively incorporates nonredundant O2M candidates with low overlap relative to O2O anchors to increase recall while maintaining near-O2O precision. We also introduce Multi-scale Adaptive Representation Diversification, which enforces structured variance and covariance constraints on PAN features, recovering 76–84% of rank lost under domain-shift compared to 26–48% without it. Across domain-shift benchmarks, RT-SFOD improves mAP by 1.4–3.5% over prior state-of-the-art methods, while the 1.3× higher throughput and smaller parameter count are inherited from the efficient YOLO backbone. We also demonstrate that our modules generalize across other dual-head detectors. Failure-case analysis and limitations are provided in the supplementary.
Acknowledgements We gratefully acknowledge the support of the Ministry of Electronics and Information Technology and the Ministry of Education, Government of India, as well as IIT Hyderabad (India) and MBZUAI (Abu Dhabi) for their support of this project. We also sincerely thank the anonymous reviewers, area chairs, and program chairs for their valuable feedback, which helped improve the quality and presentation of the paper.
16
Sairam VCR et al.
References 1. Aljalbout, E., Xing, J., Romero, A., Akinola, I., Garrett, C.R., Heiden, E., Gupta, A., Hermans, T., Narang, Y., Fox, D., et al.: The reality gap in robotics: Challenges, solutions, and best practices. Annual Review of Control, Robotics, and Autonomous Systems 9 (2025) 2. Ashraf, T., Bashir, J.: Titan: Query-token based domain adaptive adversarial learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 250–262 (2025) 3. Bardes, A., Ponce, J., LeCun, Y.: Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906 (2021) 4. Chen, C., Zheng, Z., Ding, X., Huang, Y., Dou, Q.: Harmonizing transferability and discriminability for adapting object detectors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8869–8878 (2020) 5. Chen, L., Han, J., Wang, Y.: Datr: Unsupervised domain adaptive detection transformer with dataset-level adaptation and prototypical alignment. IEEE Transactions on Image Processing 34, 982–994 (2025) 6. Chen, Q., Su, X., Zhang, X., Wang, J., Chen, J., Shen, Y., Han, C., Chen, Z., Xu, W., Li, F., et al.: Lw-detr: A transformer replacement to yolo for real-time detection. arXiv preprint arXiv:2406.03459 (2024) 7. Chen, Y., Li, W., Sakaridis, C., Dai, D., Van Gool, L.: Domain adaptive faster r-cnn for object detection in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3339–3348 (2018) 8. Chu, Q., Li, S., Chen, G., Li, K., Li, X.: Adversarial alignment for source free object detection. Proceedings of the AAAI Conference on Artificial Intelligence 37(1), 452–460 (Jun 2023) 9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213–3223 (2016) 10. Deng, J., Li, W., Chen, Y., Duan, L.: Unbiased mean teacher for cross-domain object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4091–4101 (June 2021) 11. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The international journal of robotics research 32(11), 1231–1237 (2013) 12. Hao, Y., Forest, F., Fink, O.: Simplifying source-free domain adaptation for object detection: Effective self-training strategies and performance insights. In: European Conference on Computer Vision. pp. 196–213. Springer (2024) 13. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017) 14. Huo, Y., Wu, T., Shen, Y., Li, X., Tao, Z., Yang, D.: Qrt-detr: Post-training quantization for real-time detection transformer. Neurocomputing 661, 131957 (2026) 15. Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics yolov8 (2023), https://github. com/ultralytics/ultralytics 16. Johnson-Roberson, M., Barto, C., Mehta, R., Sridhar, S.N., Rosaen, K., Vasudevan, R.: Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks? arXiv preprint arXiv:1610.01983 (2016) 17. Kennerley, M., Wang, J.G., Veeravalli, B., Tan, R.T.: Cat: Exploiting inter-class dynamics for domain adaptive object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16541–16550 (2024)
Real-Time Source-Free Object Detection
17
18. Khanam, R., Hussain, M.: Yolov11: An overview of the key architectural enhancements. arXiv preprint arXiv:2410.17725 (2024) 19. Khanh, T.L.B., Nguyen, H.H., Pham, L.H., Tran, D.N.N., Jeon, J.W.: Dynamic retraining-updating mean teacher for source-free object detection. In: European Conference on Computer Vision. pp. 328–344. Springer (2024) 20. Li, H., Zhang, R., Yao, H., Zhang, X., Hao, Y., Song, X., Peng, S., Zhao, Y., Zhao, C., Wu, Y., et al.: Seen-da: Semantic entropy guided domain-aware attention for domain adaptive object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25465–25475 (2025) 21. Li, S., Ye, M., Zhu, X., Zhou, L., Xiong, L.: Source-free object detection by learning to overlook domain style. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8014–8023 (June 2022) 22. Li, X., Chen, W., Xie, D., Yang, S., Yuan, P., Pu, S., Zhuang, Y.: A free lunch for unsupervised domain adaptive object detection without source data. Proceedings of the AAAI Conference on Artificial Intelligence 35, 8474–8481 (05 2021) 23. Li, Y.J., Dai, X., Ma, C.Y., Liu, Y.C., Chen, K., Wu, B., He, Z., Kitani, K., Vajda, P.: Cross-domain adaptive teacher for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7581–7590 (2022) 24. Liu, Q., Lin, L., Shen, Z., Yang, Z.: Periodically exchange teacher-student for source-free object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 25. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: DAB-DETR: Dynamic anchor boxes are better queries for DETR. In: International Conference on Learning Representations (2022) 26. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.E., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: Proceedings of the European Conference on Computer Vision (ECCV). Lecture Notes in Computer Science, vol. 9905, pp. 21–37. Springer (2016) 27. Ma, C., Wang, N., Zhao, Z., Chen, Q.A., Shen, C.: Slowperception: Physical-world latency attack against visual perception in autonomous driving. arXiv preprint arXiv:2406.05800 (2024) 28. Naufaldihanif, R., Kurniawan, D., Tania, K.: Performance analysis of yolo, faster r-cnn, and detr for automated personal protective equipment detection. Journal of Applied Informatics and Computing 9, 3810–3820 (12 2025) 29. Noori, M., Hakim, G.A.V., Osowiechi, D., Shakeri, F., Bahri, A., Yazdanpanah, M., Dastani, S., Ben Ayed, I., Desrosiers, C.: Histopath-c: Towards realistic domain shifts for histopathology vision-language adaptation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4890–4900 (2026) 30. RangiLyu: Nanodet-plus: Super fast and high accuracy lightweight anchor-free object detection model. (2021) 31. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence 39(6), 1137–1149 (2016) 32. Robinson, I., Robicheaux, P., Popov, M., Ramanan, D., Peri, N.: Rf-detr: Neural architecture search for real-time detection transformers (2025) 33. Safdari, R., Nikouei Mahani, M.A., Koohi-Moghadam, M., Bae, K.T.: Mixstyleflow: Domain generalization in medical image segmentation using normalizing flows. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 376–385. Springer (2025)
18
Sairam VCR et al.
34. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision 126, 973–992 (2018) 35. Sinha, S., Dwivedi, S., Azizian, M.: Towards deterministic end-to-end latency for medical ai systems in nvidia holoscan. arXiv preprint arXiv:2402.04466 (2024) 36. Sun, T., Segu, M., Postels, J., Wang, Y., Van Gool, L., Schiele, B., Tombari, F., Yu, F.: Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 21371–21382 (2022) 37. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781–10790 (2020) 38. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30 (2017) 39. Ultralytics: Ultralytics yolov5. urlhttps://github.com/ultralytics/yolov5 (Dec 2020), accessed: [Insert date here] 40. Ultralytics: YOLO26 (2025), https://github.com/ultralytics/ultralytics 41. Varailhon, S., Aminbeidokhti, M., Pedersoli, M., Granger, E.: Source-free domain adaptation for yolo object detection. In: European Conference on Computer Vision. pp. 218–235. Springer (2024) 42. VCR, S., Lalla, R., Dayal, A., Kulkarni, T., Lalla, A., Balasubramanian, V.N., Khan, M.H.: Foundation model priors enhance object focus in feature space for source-free object detection. arXiv preprint arXiv:2512.17514 (2025) 43. Vibashan, V., Oza, P., Patel, V.M.: Instance relation graph guided source-free domain adaptive object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2023) 44. Wang, A., Chen, H., Liu, L., Chen, K., Lin, Z., Han, J., et al.: Yolov10: Real-time end-to-end object detection. Advances in neural information processing systems 37, 107984–108011 (2024) 45. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 46. Wang, S., Xia, C., Lv, F., Shi, Y.: Rt-detrv3: Real-time end-to-end object detection with hierarchical dense positive supervision. In: WACV. pp. 1628–1636 (2025) 47. Wang, W., Cao, Y., Zhang, J., He, F., Zha, Z.J., Wen, Y., Tao, D.: Exploring sequence feature alignment for domain adaptive detection transformers. In: Proceedings of the 29th ACM International Conference on Multimedia. p. 1730–1738. MM ’21, Association for Computing Machinery, New York, NY, USA (2021) 48. Weng, W., Yuan, C.: Mean teacher detr with masked feature alignment: a robust domain adaptive detection transformer framework. In: Proceedings of the ThirtyEighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence. AAAI’24/IAAI’24/EAAI’24, AAAI Press (2024) 49. Wu, A., Liu, R., Han, Y., Zhu, L., Yang, Y.: Vector-decomposed disentanglement for domain-invariant object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9342–9351 (2021) 50. Yang, J., Li, C., Dai, X., Gao, J.: Focal modulation networks. Advances in Neural Information Processing Systems 35, 4203–4217 (2022)
Real-Time Source-Free Object Detection
1
51. Yao, H., Zhao, S., Li, P., Cui, Y., Lu, S., Guo, W., Lu, Y., Xu, Y., Xiong, H.: Beyond boundaries: Leveraging vision foundation models for source-free object detection. arXiv preprint arXiv:2511.07301 (2025) 52. Yao, H., Zhao, S., Lu, S., Chen, H., Li, Y., Liu, G., Xing, T., Yan, C., Tao, J., Ding, G.: Source-free object detection with detection transformer. IEEE Transactions on Image Processing 34, 5948–5963 (2025) 53. Yoon, I., Kwon, H., Kim, J., Park, J., Jang, H., Sohn, K.: Enhancing source-free domain adaptive object detection with low-confidence pseudo label distillation. In: European Conference on Computer Vision. pp. 337–353. Springer (2024) 54. Yu, F., Xian, W., Chen, Y., Liu, F., Liao, M., Madhavan, V., Darrell, T., et al.: Bdd100k: A diverse driving video database with scalable annotation tooling. arXiv preprint arXiv:1805.04687 2(5), 6 (2018) 55. Yu, J., Liu, J., Wei, X., Zhou, H., Nakata, Y., Gudovskiy, D., Okuno, T., Li, J., Keutzer, K., Zhang, S.: Mttrans: Cross-domain object detection with mean-teacher transformer. In: ECCV (2024) 56. Zhang, C.B., Zhong, Y., Han, K.: Mr. detr: Instructive multi-route training for detection transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9933–9943 (June 2025) 57. Zhang, J., Huang, J., Luo, Z., Zhang, G., Zhang, X., Lu, S.: Da-detr: Domain adaptive detection transformer with information fusion. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 23787–23798 (2021) 58. Zhao, C., Sun, Y., Wang, W., Chen, Q., Ding, E., Yang, Y., Wang, J.: Ms-detr: Efficient detr training with mixed supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17027– 17036 (June 2024) 59. Zhao, Y., Lv, W., Xu, S., Wei, J., Wang, G., Dang, Q., Liu, Y., Chen, J.: Detrs beat yolos on real-time object detection. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16965–16974 (2024) 60. Zhao, Z., Wei, S., Chen, Q., Li, D., Yang, Y., Peng, Y., Liu, Y.: Masked retraining teacher-student framework for domain adaptive object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19039–19049 (2023) 61. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)
Supplementary for Real-Time Source-Free Object Detection Sairam VCR1 , Varun Gopal1 , Poornima Jain1 , Vineeth N Balasubramanian1,2 , and Muhammad Haris Khan3 1
IIT Hyderabad, India 2 Microsoft Research 3 MBZUAI {ai20resch13001@,co22btech11015@,ai24resch11002@, vineethnb@cse.}iith.ac.in, [email protected]
Contents S.1. Feature Discriminability Analysis ................................. S.2. Qualitative Results ............................................... S.2.1. RT-SFOD is Robust to Partial Visibility ................... S.3. Pseudo-label Quality Over Training Epochs ....................... S.4. Reproducibility ................................................... S.4.1. Hyperparameter Analysis .................................. S.4.2. Dataset Details ............................................ S.4.3. Details of Data Augmentation Strategies ................... S.5. Failure-case Analysis ............................................. S.6. Teacher EMA Update Variations ................................. S.7. Analysis of MARD Module ...................................... S.8. TensorRT FP16 Efficiency Evaluation ........................... S.9. Training Overhead of Proposed Modules ......................... S.10. Additional Analyses and Clarifications ......................... S.10.1. Generality Across Dual-Head Detectors .................. S.10.2. Robustness to Cold-Start and Noisy MARD Sampling .... S.10.3. Hyperparameter Selection and Transferability ............ S.10.4. Limitations ..............................................
S.1
1 2 2 3 4 4 6 6 8 8 10 11 12 12 12 13 14 14
Feature Discriminability Analysis
Similar to Fig. 5 in the main paper, we visualize the pairwise cosine similarity (lower is better) between each anchor point’s extracted features and all other anchors across feature-scales P3/P4/P5, for two additional domain shifts: Sim10k → Cityscapes and KITTI → Cityscapes (Fig. S.1). Across domain-shifts and all feature levels, the source model exhibits highly correlated (less diverse) representations, and Mean Teacher provides only a partial reduction in similarity. In contrast, MARD consistently yields the lowest cosine similarity, indicating stronger representation discriminability and reduced feature collapse under domain-shift.
2
Sairam VCR et al.
(a) Sim10k → Cityscapes
(b) KITTI → Cityscapes
Fig. S.1: The average cosine similarity (lower is better) of each anchor point’s extracted features with all others across feature-scales (P3, P4, P5). Across domain-shifts and feature-scales, MARD consistently achieves the lowest cosine similarity.
S.2
Qualitative Results
In Fig. S.2, we provide a qualitative comparison on the Foggy Cityscapes validation set between Source-Only, Vanilla Mean-Teacher (o2o), and our RTSFOD under severe fog-induced domain shift. The Source-Only detector frequently misses objects and produces unreliable predictions, indicating poor robustness when deployed without target adaptation. While vanilla Mean-Teacher improves over the Source-Only baseline, it remains suboptimal, still exhibiting missed instances and imprecise bounding box localization. In contrast, RTSFOD produces noticeably cleaner detections by improving both coverage and quality of the predictions. Although only the one-to-one (O2O) head is used during validation, the superior predictions of RT-SFOD stem from its improved training strategy. In particular, DHF enhances the quality and completeness of pseudo-labels during training by effectively leveraging complementary supervision from dual heads, leading to more reliable target adaptation. This improved pseudo-supervision enables the student model to learn stronger representations, resulting in better object recovery and tighter localization at inference time. Furthermore, MARD mitigates the domain-shift induced degradation of multiscale feature discriminability by enforcing variance and covariance regularization on detection-aware foreground and background features across pyramid levels. Together, these components allow RT-SFOD to produce more accurate and complete detections. S.2.1
RT-SFOD is Robust to Partial Visibility
Fig. S.3 demonstrates the robustness of RT-SFOD in detecting partially visible and truncated objects in challenging foggy urban scenes. As highlighted in
Real-Time Source-Free Object Detection
3
Fig. S.2: Qualitative comparison on the Foggy Cityscapes validation set between Source-Only, Vanilla Mean-Teacher (o2o), and our proposed RT-SFOD. The red, orange, and pink arrows indicate missed detections, inaccurate localization, and incorrect class predictions, respectively. Compared to Source-Only and Vanilla Mean-Teacher, RT-SFOD successfully recovers missed objects, refines bounding box localization, and predicts the correct object categories. the examples (Fig. S.3), objects that are only marginally visible due to occlusion, truncation at image boundaries, or severe fog are still reliably detected by the model. Such scenarios are particularly demanding because only limited visual cues are available, requiring strong feature generalization and contextual understanding. The consistent detection of these partially observable objects indicates that RT-SFOD learns discriminative representations that remain effective even when object appearance is incomplete. This capability is especially critical for real-time applications such as autonomous driving, where early and reliable detection of partially visible road users can significantly enhance safety and decision-making.
S.3
Pseudo-label Quality Over Training Epochs
Fig. S.4 analyzes the evolution of pseudo-label quality produced by DHF throughout the adaptation process. We report precision, recall, and F1-score across training epochs to evaluate the reliability and completeness of the generated pseudo-labels. Precision starts at a high value and further improves steadily,
4
Sairam VCR et al.
Fig. S.3: Robust detection of partially visible objects by RT-SFOD on the Foggy Cityscapes validation set. The model successfully detects heavily occluded and truncated objects under adverse visibility conditions. The blue arrows highlight the partially visible objects. demonstrating that DHF maintains strong noise control while adaptation proceeds. More importantly, recall increases consistently over epochs, indicating that the model progressively discovers additional valid target instances as its representation becomes more domain-aligned. The resulting F1-score exhibits a smooth and monotonic upward trend, reflecting a balanced improvement in both accuracy and coverage. This behavior confirms that the proposed framework is stable during training and exhibits a self-improving mechanism.
S.4
Reproducibility
In this section, we provide the hyperparameter details of RT-SFOD, dataset descriptions, and details of the augmentations used. S.4.1
Hyperparameter analysis
We provide the hyperparameter values used by our method in Table S.1. We analyze four key hyperparameters in the main paper (Fig. 4) and three additional
Real-Time Source-Free Object Detection
5
Fig. S.4: Evolution of pseudo-label quality generated by DHF over training epochs. Precision, recall, and F1-score consistently improve as adaptation progresses, indicating stable and progressively refined pseudo supervision. ones (τno , γ, and β) in Fig. S.5. We observe that the remaining hyperparameters exhibit negligible sensitivity (≤0.3 mAP variation) over their tested ranges. As shown in Fig. S.5, performance varies by at most 0.7 mAP across all tested values of τno , γ, and β, confirming that the method does not depend on careful tuning for stable performance. Importantly, the same set of hyperparameters achieved consistent gains across all four domain-shift benchmarks (Cityscapes → Foggy Cityscapes, Cityscapes → BDD100k, Sim10k → Cityscapes, and KITTI → Cityscapes), spanning diverse shift types including weather, large-scale scene diversity, synthetic-to-real rendering gaps, and camera shift. This cross-benchmark stability indicates that RT-SFOD exhibits robustness and low sensitivity to its hyperparameter choices.
Fig. S.5: Hyperparameter sensitivity on C2F (RT-SFOD-M). mAP (%) vs. (left) the DHF non-overlap IoU threshold τno , (center ) the MARD variance target γ, and (right) the MARD covariance weight β with α fixed at 1.0. Variation is ≤0.7 mAP in all cases.
6
Sairam VCR et al.
Table S.1: Complete hyperparameter settings for RT-SFOD. The same values are used across all four domain-shift benchmarks. Sensitivity analysis plots are provided for a subset of hyperparameters; remaining ones show negligible (≤0.3 mAP) variation over their tested ranges. Group
Hyperparameter
Symbol
Value
Sensitivity
DHF
O2O confidence threshold O2M confidence threshold Non-overlap IoU threshold O2M NMS IoU threshold
τo2o τo2m τno τdup
0.5 0.5 0.2 0.7
Fig. 4 (main paper) Fig. 4 (main paper) Fig. S.5 negligible
λ0 λmax γ α β
MARD
Base regularizer weight Regularizer weight clamp Variance target Variance term weight Covariance term weight Warmup duration Confidence gate threshold Top-Kb boxes for sampling Foreground points per box Background sample points Level assignment threshold
Twarmup s̄gate Kb K M η
0.05 0.2 1.0 1.0 0.1 5 ep 0.5 15 8 128 12.0
Fig. 4 (main paper) negligible Fig. S.5 negligible Fig. S.5 negligible negligible negligible negligible negligible negligible
EMA
Teacher momentum
µ
0.999
Fig. 4 (main paper)
S.4.2
Dataset Details
We conduct experiments on five open-access benchmarks that represent four types of domain shift. The Cityscapes dataset [9] contains 5,000 high-quality annotated urban street images gathered from multiple cities and seasonal settings. From this collection, 2,925 images are used for training and 500 for validation. The dataset covers eight object classes: person, rider, car, truck, bus, train, motorcycle, and bicycle. Foggy Cityscapes [34] is derived from Cityscapes by adding simulated fog with three different density levels (0.005, 0.01, and 0.02), creating conditions that mimic reduced visibility. Following prior work [42, 12, 43], we only use the 0.02 density level images for target adaptation. Sim10k [16] consists of 10,000 synthetic urban images featuring cars, produced using the Grand Theft Auto game engine. The KITTI dataset [11] includes 7,481 annotated realworld driving images captured for autonomous driving research. Similar to existing works [43, 24, 19, 12, 41, 42, 51], for Sim10k → Cityscapes and KITTI → Cityscapes, we consider just the car class and report AP (on Car). BDD100k [54] is a large-scale driving dataset with 100,000 images collected under diverse environmental and weather conditions. S.4.3
Details of Data Augmentation Strategies
To facilitate robust self-training in our Source-Free Object Detection (SFOD) framework, we employ a Teacher-Student distillation mechanism with asymmet-
Real-Time Source-Free Object Detection
7
ric data augmentations. This section details the specific transformations applied to the target domain images. Weak Augmentation (Aweak ) The teacher model receives a weakly augmented version of the target image to ensure the stability and reliability of the generated pseudo-labels. We avoid stochastic geometric distortions here to maintain spatial consistency. The operations include: – Resize: Images are resized such that the longest edge is 1024 pixels, maintaining the original aspect ratio. – Horizontal Flip: Random horizontal flipping is applied with a probability of p = 0.5. Strong Augmentation (Astrong ) The student model is tasked with predicting the teacher’s pseudo-labels under heavy perturbation. This forces the model to rely on intrinsic object features rather than domain-specific noise or lighting conditions. The transformations include: 1. Geometric Transformations: In addition to weak augmentations, we apply random affine transformations (scaling between [0.9, 1.1] and translation up to ±5%) and perspective distortions to simulate viewpoint changes. 2. Photometric Jitter: We apply random Hue-Saturation-Value (HSV) adjustments, contrast/brightness shifts, Gamma corrections, and RGB channel shuffling to simulate varying sensor characteristics and lighting. 3. Noise and Blur: To mimic low-quality or out-of-focus captures, we inject Gaussian noise, salt-and-pepper noise, and Gaussian blur.
Table S.2: Data Augmentations used in RT-SFOD Training. Shared refers to the ones used in both weak and strong augmentations. Category
Augmentation
Range
Shared
Resize Horizontal Flip
1024 × 1024 (Long-edge) p = 0.5
Geometric
Scale/Translate Perspective
p = 0.5, scale ∈ [0.9, 1.1], trans. ±5% p = 0.3, distortion ±3%
Photometric HSV Jitter Contrast/Brightness Gamma Correction Channel Shuffle
p = 0.8, H: ±0.15, S/V: ±0.2 p = 0.6, α ∈ [0.8, 1.2], β ∈ [−20, 20] p = 0.3, γ ∈ [0.7, 1.3] p = 0.2, order ∈ {R, G, B} permutations
Degradation Gaussian Blur p = 0.4, kernel ∈ {3, 5, 7}, σ ∈ [0.5, 2.0] Noise (Gaussian) p = 0.3, σstd ∈ [5, 15] Noise (Salt & Pepper) p = 0.3, prob ∈ [0.01, 0.05]
8
Sairam VCR et al.
Fig. S.6: Failure cases of RT-SFOD on the Foggy Cityscapes validation set under extremely challenging conditions. Typical errors occur in crowded smallobject scenarios and under severe fog and overexposure, where visibility is significantly degraded.
S.5
Failure-case Analysis
Fig. S.6 illustrates challenging scenarios where RT-SFOD encounters limitations under extreme visual degradation. In the first row, densely packed and small objects, such as bicycles, create heavy occlusion and overlap. In such highly crowded regions, the detector may miss a few instances or produce a single bounding box covering multiple objects due to the limited spatial separability at that scale. The second row presents cases with extremely dense fog and strong sunlight exposure, where object boundaries are severely washed out, and contrast is drastically reduced, sometimes making objects barely distinguishable even to the human eye. Under these rare but severe conditions, RT-SFOD may produce suboptimal localization. Nevertheless, these examples highlight the inherent difficulty of the scenarios rather than the systematic weaknesses of the method. Overall, RT-SFOD remains robust across the vast majority of targetdomain images and demonstrates strong resilience to domain shift, with failures primarily occurring in extreme edge cases involving severe visibility degradation and dense small-object clustering.
S.6
Teacher EMA Update Variations
A known failure mode of fully-unlabeled teacher-student self-training is catastrophic collapse, where degraded teacher predictions are amplified through the pseudo-label feedback loop and propagate to the student. PETS [24] mitigates this via stronger temporal separation (periodic teacher–student exchange and an
Real-Time Source-Free Object Detection
9
Fig. S.7: Impact of EMA teacher update frequency (momentum = 0.999) on target-domain performance (mAP%) - C2F domain-shift. Updating the teacher once per epoch achieves the best performance (47.4%), outperforming periteration updates, less frequent updates, and a frozen teacher (no update).
additional slower teacher), highlighting the importance of keeping the teacher as a stable target rather than a rapidly moving copy of the student. Simple-SFOD [12] shows that Adaptive Batch Normalization (AdaBN) provides a better starting point on the target domain and makes fixed pseudo-label training competitive with EMA-based methods. However, AdaBN does not eliminate pseudo-label noise or the non-stationarity introduced by augmentation and thresholding.
Figure S.7 studies the effect of teacher update frequency under a fixed EMA momentum of 0.999 on our RT-SFOD-M. Updating the teacher at every iteration results in lower performance (44.0%), indicating that overly frequent updates cause the teacher to closely track short-term student fluctuations, which may amplify noisy pseudo-labels. In contrast, a frozen teacher (no update) achieves 46.2%, suggesting that a stable target is beneficial but limited by its inability to improve pseudo-label quality over time. Interestingly, updating every half epoch (45.8%) underperforms a frozen teacher (46.2%), indicating that intermediate update frequencies may introduce pseudo-label drift without sufficient temporal separation to stabilize training. Updating the teacher once per epoch achieves the best performance (47.4%), outperforming both more frequent and less frequent updates. This schedule maintains sufficient temporal separation to avoid reinforcing transient student errors, while still allowing the teacher to progressively incorporate improved target-adapted representations. Updating every two epochs (46.5%) slightly degrades performance, indicating that infrequent updates reduce the benefits of teacher adaptation.
10
Sairam VCR et al.
(a) Domain-shift reduces feature effective rank
(b) Rank recovery comparison
Fig. S.8: Analysis of effective-rank degradation and recovery on Cityscapes→Foggy Cityscapes. (a) Domain shift reduces effective rank across feature scales. (b) A plot of Normalized rank recovery (Eq. S.3) shows that MARD restores substantially more multi-scale feature diversity than vanilla mean-teacher self-training, closing the gap toward the target oracle.
S.7
Analysis of MARD Module
In the MARD module, given n channel-wise normalized tokens comprising Z̃ℓ , we define the covariance matrix for each level ℓ as: \text {Cov}(\tilde {Z}_\ell ) = \frac {1}{n-1} \sum _{i=1}^{n} (\tilde {z}_i - \bar {\tilde {z}})(\tilde {z}_i - \bar {\tilde {z}})^T, \quad \text {where} \quad \bar {\tilde {z}} = \frac {1}{n} \sum _{i=1}^{n} \tilde {z}_i.
(S.1)
Here, Cov(Z̃ℓ ) represents the sample covariance matrix of the channel-wise normalized feature-vectors at the ℓ-th level. The variable n denotes the total number of feature-vectors, while z̃i corresponds to the i-th individual normalized featurevector. The term z̃¯ is the empirical mean of all n feature-vectors within this level. ¯ i − z̃) ¯ T captures the pairwise feature correlations The outer product (z̃i − z̃)(z̃ 1 across the channels, and scaling by n−1 provides an unbiased estimate of the covariance matrix. We define the channel wise variance as: \text {Var}(Z_{\ell ,:,j}) = \frac {1}{n-1} \sum _{i=1}^{n} (z_{i,j} - \bar {z}_j)^2, \quad \text {where} \quad \bar {z}_j = \frac {1}{n} \sum _{i=1}^{n} z_{i,j}.
(S.2)
In this formulation, Var(Zℓ )j denotes the sample variance of the j-th channel across all n feature-vectors at level ℓ. The term zi,j represents the activation of the j-th channel for the i-th feature-vector, and z̄j is the empirical mean of that specific channel over all feature-vectors. In Fig. 2(b) (main paper), we quantify multi-scale feature discriminability using the effective rank of the YOLOv10 PAN feature maps at levels
Real-Time Source-Free Object Detection
11
ℓ ∈ {P 3, P 4, P 5}. We provide further supplemental analysis of MARD here. Similar to Fig. 2(b), Source Oracle, refers to the source-trained model evaluated on the source validation set (Cityscapes→Cityscapes); No Adaptation refers to the source-trained model evaluated directly on the target validation set (Cityscapes→Foggy Cityscapes); and the Target Oracle refers to a fully supervised model trained and evaluated on the target domain (Foggy Cityscapes→Foggy Cityscapes). Fig. S.8a shows the phenomenon of a consistent drop in effective rank induced due to domain shift, across feature scales. We define a normalized rank recovery score at each scale ℓ: \mathrm {Recovery}_\ell = \frac {\mathrm {Rank}_\ell - \mathrm {NoAdapt}_\ell }{\mathrm {TargetOracle}_\ell - \mathrm {NoAdapt}_\ell }, \label {eq:rank_recovery}
(S.3)
where Rankℓ denotes the effective rank obtained by a given adaptation variant at scale ℓ, NoAdaptℓ is the effective rank of the source model on the target domain (no adaptation), and TargetOracleℓ is the effective rank of a target-supervised oracle on the target domain. This normalization expresses recovery as the fraction of the recoverable gap from NoAdaptℓ to the target oracle: Recoveryℓ = 0 corresponds to no improvement over NoAdaptℓ , and Recoveryℓ = 1 corresponds to fully closing the gap to the target oracle. Fig. S.8b plots Recoveryℓ for vanilla mean-teacher adaptation and for our proposed MARD. Consistent with Fig. 2(b), vanilla self-training recovers only a limited portion of rank, whereas MARD recovers substantially more rank at all pyramid levels, demonstrating that it effectively counters the domain-shift-induced feature collapse, hence preserving multi-scale feature discriminability.
S.8
TensorRT FP16 Efficiency Evaluation
As discussed in the main paper (Sec. 4), prior efficiency-focused object detection works [59, 41, 44] frequently report latency using TensorRT FP16, while reporting accuracy in FP32 PyTorch. To facilitate reproducibility and provide a reference under this commonly reported deployment setting, we additionally report TensorRT FP16 latency for RT-SFOD in Table S.3. These results are provided for completeness and comparability only. All primary comparisons in the main paper remain based on FP32 PyTorch evaluation. Experimental Protocol. We measure latency using TensorRT FP16 on a Tesla T4 GPU. Models are exported with an input resolution of 1 × 3 × 512 × 1024 and batch size 1. Timing is performed using CUDA events after a warmup phase of 100 iterations, followed by 500 measured iterations. We report forward-pass execution time only, excluding image decoding, preprocessing, host-device transfer, and post-processing. The statistics p50, p90, and p99 denote the 50th, 90th, and 99th percentile latency values across measured iterations, respectively. These numbers reflect the efficiency of RT-SFOD, which reduces to a standard YOLOv10 detector during inference. The reported latency is not directly comparable to the FP32 PyTorch efficiency values in the main paper due to differences in precision, runtime framework, hardware, and measurement protocol.
12
Sairam VCR et al.
Table S.3: TensorRT FP16 forward-pass latency on a Tesla T4 GPU, batch size 1, input resolution 512 × 1024. Model RT-SFOD-S RT-SFOD-M RT-SFOD-L
S.9
Mean (ms) p50 3.611 7.795 12.253
p90
p99
FPS
3.633 3.808 6.559 276.93 7.769 7.926 8.286 128.29 12.187 12.558 12.745 81.61
Training Overhead of Proposed Modules
Table S.4: Training overhead of proposed modules relative to Vanilla MeanTeacher (RT-SFOD-M, Cityscapes → Foggy Cityscapes, single NVIDIA RTX A6000 GPU, batch size 16). Configuration Vanilla MT (baseline) + DHF + MARD + DHF + MARD (RT-SFOD)
Peak Mem. (MB) ∆ Mem. Time/Epoch (s) 21,120 21,400 22,360 22,590
— +280 +1,240 +1,470
139.0 145.4 156.0 162.7
∆ Time — +6.4 s (+4.6%) +17.0 s (+12.2%) +23.7 s (+17.1%)
Table. S.4 reports the peak GPU memory and time per epoch for each configuration, measured on a single NVIDIA RTX A6000 GPU with batch size 16 on the C2F benchmark. Both DHF and MARD are training-time only modules that do not affect the deployed model, so this overhead has no bearing on inference FPS or latency reported in the main paper. DHF introduces negligible cost: +280 MB (+1.3%) in peak memory and +6.4 s (+4.6%) per epoch, with no architectural changes to the student or teacher networks. The slightly higher overhead of MARD (+1,240 MB, +17.0 s over baseline) stems from computing per-level variance and covariance statistics over foreground and background feature vectors sampled from the multiscale PAN outputs. The full RT-SFOD configuration (DHF + MARD) incurs a combined overhead of +1,470 MB and +23.7 s (+17.1%) per epoch over Vanilla Mean-Teacher, a modest training-time cost in exchange for consistent 1.4 to 3.5% mAP gains across four benchmarks. The overhead is consistent across all domain-shift benchmarks.
S.10
Additional Analyses and Clarifications
This section consolidates additional experiments and clarifications that further support the claims made in the main paper. S.10.1
Generality Across Dual-Head Detectors
RT-SFOD targets NMS-free dual-head detectors in general, not YOLOv10 specifically (Contribution 1, main paper). DHF requires only a precise one-to-one
Real-Time Source-Free Object Detection
13
(O2O) head and a dense, high-recall one-to-many (O2M) head—an increasingly common design that combines strong supervision with NMS-free efficient inference. Representative examples include H-DETR (CVPR’23), MS-DETR (CVPR’24), RT-DETRv3 (WACV’25), Mr. DETR (CVPR’25), and YOLOv26. To verify that our findings are not tied to a single architecture, we apply RTSFOD with the same hyperparameters to three additional dual-head detectors on the Cityscapes → Foggy Cityscapes (C2F) benchmark. Table S.5: Additional C2F results vs. the MT(O2O) baseline on YOLOv26S/M/L, MS-DETR (CVPR’24), and Mr. DETR (CVPR’25), using the same hyperparameters as YOLOv10. Detector YOLOv26S YOLOv26M YOLOv26L MS-DETR Mr. DETR
MT (O2O) RT-SFOD 42.1 46.0 50.6 42.3 43.8
44.6 (+2.5) 48.8 (+2.8) 52.7 (+2.1) 44.3 (+2.0) 46.1 (+2.3)
As shown in Table S.5, RT-SFOD improves the MT (O2O) baseline on YOLOv26S/M/L by +2.5/+2.8/+2.1 mAP, on MS-DETR by +2.0, and on Mr. DETR by +2.3, consistent with the gains observed on YOLOv10. This indicates that the precision–recall fusion exploited by DHF and the featurediversification effect of MARD are properties of the dual-head paradigm rather than of one specific detector. For single-head detectors (e.g., YOLOv5/v11), DHF has no second candidate set and gracefully reduces to confidence-threshold pseudo-labeling as in SF-YOLO; MARD remains applicable whenever pseudolabels and multi-scale features are available.
S.10.2
Robustness to Cold-Start and Noisy MARD Sampling
A natural concern with self-training under large domain gaps is error propagation: if the teacher’s initial predictions are unreliable, the DHF anchors and the regions MARD samples could be corrupted from the outset. Cold start. AdaBN warm-starts the teacher before self-training begins, and Fig. S.4 shows that DHF pseudo-label precision is already 0.920 at epoch 0, so the O2O anchors are reliable before adaptation. To probe a substantially larger initial gap than our driving benchmarks, we evaluate RT-SFOD-M on the extreme PASCAL VOC → Clipart shift, where it improves the MT (O2O) baseline from 37.8 to 40.5 mAP. Noisy MARD sampling. MARD mitigates noisy supervision through warmstarting, confidence-gating, and periodic application. To stress-test its sensitivity to inaccurate sampling regions directly, we corrupt the initial pseudo boxes on C2F (RT-SFOD-M) by jittering their centers and sizes for 20%/40%/60%
14
Sairam VCR et al.
of boxes. mAP remains stable, changing only from 47.4 to 47.2/47.0/46.9, respectively. Together, these results show that RT-SFOD is robust both to severe cold-start shifts and to inaccurate early MARD sampling. S.10.3
Hyperparameter Selection and Transferability
MARD sampling counts. The foreground/background sampling counts K/M control how many feature vectors are drawn per pseudo box; the aim is to cover enough object and background regions without noisy over-sampling or excessive computational overhead. We swept K ∈ {2, 4, 8, 16, 32} and M ∈ {16, 32, 64, 128, 256} and found K=8, M =128 to give the best mAP, with negligible sensitivity (≤0.3 mAP) over the tested ranges (Table S.1). Transfer across scales and architectures. We tuned hyperparameters once on YOLOv10-M and reused the same values across all model scales and the additional dual-head detectors in Sec. S.10.1, where they yield consistent gains. For example, varying τo2o over {0.3, 0.4, 0.5, 0.6, 0.7} on YOLOv26S changes mAP by ≤1%. We attribute this transferability to the dependence of our thresholds on relative statistics rather than architecture-specific factors: confidence is internally calibrated (a threshold of 0.5 still selects reliable predictions in each detector), IoU is purely geometric, and domain shift affects multi-scale features similarly across detectors. As a result, a single shared configuration continues to work well across models and scales. S.10.4
Limitations
While RT-SFOD advances the accuracy–speed–size trade-off for source-free detection, it has a few limitations. (i) Architecture scope. DHF is designed for NMS-free dual-head detectors; on single-head detectors it has no second candidate set and degenerates to simple confidence-threshold pseudo-labeling (as in SF-YOLO), so the pseudo-label fusion gains reported here are specific to the dual-head setting. MARD, however, remains applicable whenever pseudolabels and multi-scale features are available. (ii) Capacity on camera-shift. On camera-shift benchmarks such as K2C, performance benefits from additional model capacity, as discussed in Sec. 4.1. (iii) Extreme visual degradation. As shown in Fig. S.6, residual failures occur under densely packed small objects and severe fog or overexposure, where visual cues are minimal. These cases reflect the inherent difficulty of the inputs rather than systematic weaknesses of the method, and RT-SFOD remains robust across the vast majority of targetdomain images.