Budget-Adaptive Routing: Skipping the Weak When the Strong Answers Anyway Wei Geng
Technical University of Munich Munich, Germany [email protected]
Nitinder Mohan
TU Delft Delft, Netherlands [email protected]
Jörg Ott
Technical University of Munich Munich, Germany [email protected]
arXiv:2606.30919v1 [cs.NI] 29 Jun 2026
Abstract Edge-cloud inference collaborations are often designed with a routing estimator1 that decides whether to offload each frame from weak models at the edge to stronger models in the cloud. Existing systems place the routing estimator after the weak detector, so the weak forward pass still runs even on frames that are later offloaded. In this paper, we argue that this weak-conditioned design can be suboptimal when the offload budget varies. First, we present a competitive weak-skipping estimator (0.153 GFLOPs, ∼29× lighter than the weak detector at 4.49 GFLOPs) that extracts routing signal from raw pixels, outperforming the common after-weak placement weak-conditioned baselines. Second, we show that neither weakskipping nor weak-conditioned placement dominates across the full operating curve, and we propose budget-adaptive routing, which selects between them by offload budget via two offline-tuned thresholds. On PASCAL VOC, our budget-adaptive router traces the upper accuracy envelope of both fixed placements across the operating range. Our method 2 reduces per-frame latency by up to 19.1 ms (∼30% lower at 𝜌=0.9). Besides outperforming SOTA methods, it is surprisingly stronger than the strong model (+1.7 pp over the strong model’s peak mAP) at some operating points with far less compute.
CCS Concepts • Networks → Cloud computing; Network performance modeling; • Computer systems organization → n-tier architectures; • Computing methodologies → Object detection.
Keywords selective offloading, budget-adaptive routing, object detection, costaccuracy trade-off ACM Reference Format: Wei Geng, Nitinder Mohan, and Jörg Ott. 2026. Budget-Adaptive Routing: Skipping the Weak When the Strong Answers Anyway. In Workshop on Networks for AI Computing (SIGCOMM ’26), August 17–21, 2026, Denver, CO, USA. ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/3789240. 3828740 1 In this paper, in most cases we use estimator and router interchangeably. 2 Artifacts are available at § https://github.com/ViGeng/bgt-ada
This work is licensed under a Creative Commons Attribution 4.0 International License. SIGCOMM ’26, Denver, CO, USA © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2467-1/26/08 https://doi.org/10.1145/3789240.3828740
Figure 1: Selective offloading for object detection: a local weak detector and a cloud strong detector with a per-frame router under offload budget 𝜌.
1
Introduction
Edge devices increasingly run visual perception pipelines by offloading computation to the cloud, such as traffic cameras counting vehicles, industrial robots operating on assembly lines, and identity systems checking faces at borders. While some simply stream tasks to the edge or cloud for inference [2, 5, 8, 13, 18], others pair a local/edge weak detector that is fast but limited with a strong detector in the cloud that is more accurate but more costly in latency, compute, bandwidth, and energy. Selective offloading routes each frame so that the cloud complements local inference, reducing overall cost without sacrificing, and often improving, end-to-end accuracy [3, 23]. This routing decision is constrained by an offload budget that caps the fraction of frames sent to the cloud, given network conditions, compute resources, and application requirements. A decade of selective-offloading work [3, 8, 12, 15, 23, 26, 28] runs the weak detector on every frame and feeds its output (proposal scores, top-𝑘 box statistics, learned embeddings, or intermediate activations) to a routing estimator that decides whether to escalate to a stronger detector in the cloud. EdgeML [23] regresses on the top-25 proposal features so that uncertain cases can be improved by strong models. DCSB [3] hand-crafts a difficult-case discriminator that directs requests to the strong model at a fixed threshold. Earlyexit families such as BranchyNet [26], MSDNet [12], and SDN [15] gate within the weak network at intermediate layers. Despite their architectural variety, all of these methods share one structural commitment: the weak forward pass runs on every frame, because the routing decision depends on its output or intermediate features.
SIGCOMM ’26, August 17–21, 2026, Denver, CO, USA
2 Problem Statement 2.1 Problem Formulation Let I = (𝑖 1, . . . , 𝑖 𝑁 ) be a stream of image frames, 𝑀𝑤 a weak local detector with per-frame compute cost 𝐶 𝑤 , and 𝑀𝑠 a strong cloud detector with per-frame cost 𝐶𝑠 (compute plus network round-trip). For each frame 𝑖𝑡 the router produces a binary routing decision 𝑑𝑡 ∈ {0, 1} where 𝑑𝑡 = 0 means “return 𝑀𝑤 (𝑖𝑡 )” and 𝑑𝑡 = 1 means
Conditioned
Skipping (ours)
Partial Model
Strong Model
Strong Model bypass
Partitioned
suspicious
initial
Weak Model
Cloud Edge
Weak Model
Estimator results
Partial Model
results
intermediate results
This commitment is misaligned with high-budget deployments. Safety-critical settings such as autonomous driving or security run at a high offload budget, say ≥ 30% of frames sent to the cloud, and there the weak forward pass is wasted on most frames: they are offloaded and answered by the strong model anyway, so paying for the weak pass only adds unnecessary compute and latency. Structurally, if the estimator is moved before the weak model and predicts from the raw image, the serial dependency on the weak model is removed and the weak pass can be skipped for frames that will most likely be offloaded. We call this a weak-skipping estimator, in contrast to the weak-conditioned estimators of prior work (shown in Fig. 2 left and middle). To our knowledge no previous selectiveoffloading system has deployed it. The usual assumption is that rawpixel features are too weak to predict detector failure. Our results challenge this: a 0.15 GFLOPs image-only weak-skipping estimator trained on a binary offload-utility target (Eq. (6)) is competitive with proposal-feature baselines that require the full 4.49 GFLOPs weak forward pass. We attribute this to the assumption that predicting whether a frame is difficult is substantially cheaper than predicting what it contains. Weak-skipping is not, however, strictly better. At low offload budgets the weak forward pass is unavoidable anyway. Since the cloud rarely fires, the weak-conditioned estimator’s richer features come essentially for free. The two placements therefore divide the operating curve into compute regimes: weak-conditioned and weakskipping win in the low and high-budget bands, respectively, and neither dominates the curve. Taken together, these observations argue for an adaptive router that selects between weak-skipping and weak-conditioned placements as a function of the offload budget, instead of one fixed placement. We confirm it empirically on PASCAL VOC from compute, latency, and accuracy perspectives. This paper contributes: (i) We articulate the weak-first assumption latent in the selectiveoffloading literature, formalize its hidden cost as an implicit compute tax. (§ 2). (ii) We introduce the first competitive weak-skipping estimator for object-detection offloading: a 0.15 GFLOPs image-only estimator that matches or exceeds the strongest weak-conditioned baselines (§ 3.1) detection quality-wise. We also build a lightweight weak-conditioned estimator (XGBoost on MORIC) that outperforms current weak-conditioned SOTAs on routing quality at sub-ms inference cost.(Tab. 2). (iii) Building on our weak-skipping and weak-conditioned estimators, we propose budget-adaptive routing, which selects between weak-skipping and weak-conditioned placements according to the deployment budget. From our simulated study, our approach outperforms SOTAs. (§ 3.2, § 4).
Geng et al.
easy
Estimator
Figure 2: Routing schemes: full partitioned compute [12, 14, 15, 26] (left), weak-conditioned after the weak pass [3, 23] (middle), and our weak-skipping (right), which enables budget-adaptive routing. Estimator placement is the axis: it fires after the weak model (middle) or before it (right). Green marks the keep-local path, red the escalation to the cloud. “return 𝑀𝑠 (𝑖𝑡 )”. Given a downstream detection utility 𝑈 (·) (typically mean Average Precision, mAP), an offload budget 𝜌 ∈ (0, 1], and an estimator producing a per-frame score 𝑠𝑡 , the router solves 1 ∑︁ 𝑈 𝑖𝑡 , 𝑀𝑑𝑡 𝑁 𝑡 =1 𝑁
max 𝑑 1:𝑁
1 ∑︁ 𝑑𝑡 ≤ 𝜌. 𝑁 𝑡 =1 𝑁
s.t.
(1)
We instantiate 𝑈 as [email protected], the standard PASCAL VOC metric [6], for comparability with the VOC benchmark and the weakconditioned baselines we reproduce. The framework is otherwise metric-agnostic: the proxy reward ΔAP (Eq. (4)) can be computed for any 𝑈 , so a stricter COCO-style AP@[.5:.95] is a drop-in substitution we leave to future work.
2.2
The Implicit Compute Tax
In every weak-conditioned design the weak pass executes on every frame, including those answered by the cloud, as an implicit compute tax. The expected per-frame compute is
𝐶𝑒cond
𝑇cond (𝜌) = 𝐶 𝑤 + 𝐶𝑒cond + 𝜌 · 𝐶𝑠 ,
(2)
skip 𝑇skip (𝜌) = 𝐶𝑒 + (1 − 𝜌) · 𝐶 𝑤 + 𝜌 · 𝐶𝑠 ,
(3)
skip 𝐶𝑒
where and are the estimator costs for the two skip placements. The tax is 𝜌 𝐶 𝑤 − (𝐶𝑒 − 𝐶𝑒cond ). With our values (MobileNetV3 [11] vs ResNet50 [10], 𝐶 𝑤 =4.49, 𝐶𝑒cond ≈0, skip 𝐶𝑒 =0.15, 𝐶𝑠 =280.37 GFLOPs), the tax reaches 3.90 GFLOPs at 𝜌=0.9, approaching the full weak-detector cost. In addition, a weak-conditioned router cannot start until the weak forward pass finishes (24.70 ms here), so 𝐶 𝑤 is a serial wall-clock dependency on every frame. The tax reflects a fallback mindset: the cloud as a backstop for the weak detector. Beside the compute and latency costs, it may also increase the jitter of the system (we leave further analysis to a follow-up work), which is undesirable for real-time applications. We don’t have to pay the tax if we treat the cloud as a collaborator that complements local inference instead of a fallback, illustrated as a branch topology in Fig. 2-right. Then the practical question
Budget-Adaptive Routing: Skipping the Weak When the Strong Answers Anyway
SIGCOMM ’26, August 17–21, 2026, Denver, CO, USA
is whether an image-only lightweight estimator can catch enough routing signal without paying the tax?
62.6% @ 0
where 𝐹 + and 𝐹 − are the empirical CDFs of ΔAP over the frames where offloading strictly helps (ΔAP>0) and strictly hurts (ΔAP<0), respectively; MORIC+ (𝑖𝑡 ) ∈ [−1, 1] and OffloadBin(𝑖𝑡 ) ∈ {0, 1}. MORIC+ generalises EdgeML’s MORIC [23] by splitting the CDF at zero, which equalises positive and negative magnitudes and gives a symmetric loss landscape around the routing boundary. OffloadBin instead reduces the problem to a binary class label, which we train with focal loss [20] to handle the positive-class imbalance. On VOC the raw ΔAP is long-tailed with 62.6% of frames at exactly zero, 25.1% positive, and 12.3% negative (Fig. 3), so direct regression wastes capacity on the zero spike. In other words, the estimator is always lazily and blindly predicting the majority class zero, which already yields good rewards. As a result, it fails to learn the routing boundary. Both transformed targets decouple hardness from content, which is what lets raw pixels suffice. OffloadBin gives our best routing quality and MORIC+ is the softer-signal variant reported in the per-ratio sweep. Architecture and calibration. A weak-skipping estimator 𝑓skip : 𝐼 ↦→ 𝑠ˆskip produces a routing score from the raw image alone, where 𝑠ˆskip ∈ [0, 1] estimates 𝑃 (ΔAP>0) when trained on OffloadBin (or the predicted MORIC+ ∈ [−1, 1]); a higher score means offloading is more likely to help (not how much to help), instantiating the generic per-frame score 𝑠𝑡 of § 2.1. Our reference instantiation is a highly compressed MobileNetV2-Lite [25] backbone (128×128 input, 0.15 GFLOPs, 0.54 M parameters) trained on OffloadBin, though the framework is agnostic to the backbone and proxy. A thresholder 𝜋 𝜌 [9] then maps the score stream to binary decisions 𝑑𝑡 =𝜋 𝜌 (𝑠ˆskip )
CDF
MORIC+
Proxy metric. A weak-skipping estimator can only out-cheap a weak-conditioned one if it learns to predict whether offloading helps, not the contents of the frame. A naive per-frame target like ΔAP (= cloud AP − local AP), is misleading because detection AP is computed dataset-wide via a single global Precision-Recall (PR) curve, so a single frame’s contribution depends on every other frame. Inspired by EdgeML [23], we use the contextual offloading reward: for frame 𝑖𝑡 , the per-frame reward ΔAP(𝑖𝑡 ) = AP swap𝑡 → 𝑠 − AP all-weak , (4)
74.9%
OffloadBin
i.e., the change in dataset-wide [email protected] obtained by replacing the local detections on 𝑖𝑡 with the cloud detections while holding all other frames at their local outputs. Computing it offline over the training set is O (𝑁 ) via a precomputed-IoU merge swap, and the resulting reward depends only on the dataset, not on the system at inference time. Eq. (4) produces a long-tailed signed signal that is hard to regress directly, as shown by the raw ΔAP ridge (top) of Fig. 3. We extract two estimator-friendly targets: + 𝐹 ΔAP ΔAP(𝑖𝑡 ) ΔAP(𝑖𝑡 ) > 0, MORIC+ (𝑖𝑡 ) = 0 (5) ΔAP(𝑖𝑡 ) = 0, 𝐹 − ΔAP(𝑖𝑡 ) − 1 ΔAP(𝑖𝑡 ) < 0, ΔAP OffloadBin(𝑖𝑡 ) = 1 ΔAP(𝑖𝑡 ) > 0 , (6)
CDF
raw ΔAP
3 Method 3.1 Weak-skipping Estimator
CDF
25.1%
−1.0
−0.5
0.0
0.5
1.0
learning target value
Figure 3: Per-frame learning targets on VOC test (𝑁 =3105). Top: raw ΔAP is degenerate, with 62.6% of frames exactly at 0 (stem) and short signed tails (25.1%>0, 12.3%<0). Middle: MORIC+ spreads this into a smooth, symmetric regression target on [−1, 1]. Bottom: OffloadBin (1[ΔAP>0]) is the binary classification target, a ∼1:3 split (25.1% positive). so that the offloaded fraction matches the requested budget 𝜌 without test-set lookahead. Concretely, 𝜋 𝜌 keeps a running estimate of the (1−𝜌) quantile of the incoming scores and offloads any frame scoring above it, nudging the threshold as the stream drifts so the realized offload rate stays near 𝜌. Ratio control thus stays orthogonal to the score: any estimator emitting a calibrated per-frame score drops into the same thresholder. We leave a full treatment of the thresholder, including drift handling, finite-window error, and the resulting jitter, to follow-up work.
3.2
Budget-adaptive Routing
The accuracy side of the placement mirrors the compute side of § 2.2. At low budgets, weak-skipping wastes richer signals from the weak forward pass on most frames because most of them comes for free. At high budgets, weak-conditioned pays the tax on most frames because they will be offloaded anyway. Every fixed-placement router is therefore suboptimal on part of the operating curve. Formulation. A budget-adaptive router holds two estimators, 𝑓skip : 𝐼 ↦→ 𝑠ˆskip and 𝑓cond : (𝐼, 𝑀𝑤 (𝐼 )) ↦→ 𝑠ˆcond (ˆ𝑠 cond ∈ [0, 1], read from the image plus weak output), plus a binary arbiter 𝛼 (𝜌) ∈ {0, 1} and the same thresholder 𝜋 𝜌 (§ 3.1), now shared across both estimators. The per-frame decision is 𝑑𝑡 = 𝜋 𝜌 𝛼 (𝜌) 𝑠ˆskip (𝑖𝑡 ) + (1−𝛼 (𝜌)) 𝑠ˆcond (𝑖𝑡 , 𝑀𝑤 (𝑖𝑡 )) . (7) c 𝜌, 𝑎 𝑓skip + (1−𝑎) 𝑓cond , 𝛼 (𝜌) = arg max AP (8) 𝑎∈ {0,1}
c is offline [email protected] on a held-out tuning split. Evaluated per where AP budget, Eq. (8) makes 𝛼 (𝜌) piecewise constant with two crossovers,
SIGCOMM ’26, August 17–21, 2026, Denver, CO, USA
splitting the operating range into three regimes: 0 𝜌 < 𝜌 frontier (weak-conditioned), 𝛼 (𝜌) = 1 𝜌 frontier ≤ 𝜌 < 𝜌 ceiling (weak-skipping), 0 𝜌 ≥ 𝜌 ceiling (weak-conditioned).
Geng et al.
(9)
The two thresholds are the budgets at which the winning placement flips on the tuning split. At the low crossover 𝜌 frontier the weak pass on offloaded frames turns from a near-free byproduct into pure tax, so weak-skipping starts to win. At the high crossover 𝜌 ceiling the fewer frames still kept local carry enough weight that the richer weak-conditioned signal wins again. We fit both once, offline, by sweeping 𝜌 and reading off the two switch points (on VOC, 𝜌 frontier =0.3 and 𝜌 ceiling =0.8). At runtime the arbiter is a constanttime lookup on 𝜌.
4
Preliminary Evidence
We report preliminary results on PASCAL VOC [6] based on the setup in Tab. 1: the weak/strong mAP gap is small which makes routing genuinely difficult and any larger gap can make the benefits more pronounced. Three claims are verified: (i) Weak-skipping estimators are competitive with weakconditioned baselines on routing quality, and our weakconditioned estimator outperforms current SOTAs (§ 4.1, Tabs. 2 and 3, Fig. 5). (ii) The compute advantage of weak-skipping routing at high offload budget is real GFLOPs- and latency-wise. (§ 4.2, Fig. 4). (iii) A budget-adaptive router that selects between the two placements traces the lowest cost across the operating curve and Pareto-dominates either fixed placement on the joint accuracycompute frontier, outperforming all existing baselines (§ 4.3, Fig. 5 and Tab. 3). Table 1: Experiment setup (PASCAL VOC). Detectors (profiled on Nvidia A40) Weak 𝑀𝑤 fasterrcnn_mobilenet_v3_large_fpn [11, 19, 24]; 𝐶 𝑤 =4.49 GFLOPs / 24.70 ms; mAP = 0.760 Strong 𝑀𝑠 fasterrcnn_resnet50_fpn_v2 [10, 19, 24]; 𝐶𝑠 =280.37 GFLOPs / 44.12 ms; mAP = 0.791 Estimators (ours, § 3) WeakMobileNetV2-Lite [25] on OffloadBin or MORIC+ skipping WeakXGBoost [4] on MORIC conditioned Budgetoffline-tuned 𝜌 frontier , 𝜌 ceiling adaptive Prior weak-conditioned baselines EdgeML [23] proposal-level, after weak detector DCSB [3] fixed-rule, after weak detector Reference points Trivial always-weak, always-strong, uniformly random Oracle per-frame ΔAP ground-truth gain
4.1
Weak-skipping is Empirically Viable
Tab. 2 reports an overview across all estimators on VOC Test. Our weak-skipping MobileNetV2-Lite estimator trained on OffloadBin attains a Spearman rank correlation of 0.557 with the per-frame
Table 2: Routing quality on VOC (single-pass, seed 42, mAP at IoU 0.5). Spearman 𝜌𝑠 vs. the per-frame oracle; Peak mAP and AUC𝜌 (area under the mAP–𝜌 curve) over 𝜌 ∈ [0, 1] with step 0.1. Green deltas on Peak mAP are relative % over always-weak (0.760). weak-conditioned methods additionally pay 𝐶 𝑤 =4.49 GFLOPs upstream. “–” = no continuous score. Estimator
𝜌𝑠
Peak mAP
AUC𝜌
GFLOPs
Reference points Always weak Always strong Uniform random Oracle (ΔAP)
0.000 0.000 0.000 –
0.760 0.00%= 0.791 4.08%▲ – 0.827 8.82%▲
0.760 0.791 0.777 0.816
0 0 0 –
Weak-conditioned (after weak) EdgeML [23] DCSB [3] XGBoost [4] / MORIC (Ours)
0.138 0.359 0.472
0.793 4.34%▲ 0.789 3.82%▲ 0.804 5.79%▲
0.784 – 0.794
≈0 ≈0 ≈0
Weak-skipping (before weak) MV2-Lite + MORIC+ (Ours) MV2-Lite + OffloadBin (Ours)
0.350 0.801 5.39%▲ 0.557 0.808 6.32%▲
0.790 0.795
0.15 0.15
Budget-adaptive (𝜌 frontier =0.3, 𝜌 ceiling =0.8) OffloadBin ↔ MORIC (Ours) –‡ 0.808 6.32%▲
0.796
0.15†
† Adaptive uses 𝑓 cond for 𝜌 ≤0.2 and 𝜌 ≥0.8, 𝑓skip for 0.3≤𝜌 ≤0.7 (offline arbitration, Eq. (8)).
Reported GFLOPs are the skipping-branch cost; the weak-conditioned branch additionally pays 𝐶 𝑤 . ‡ Spearman is undefined for an arbiter that selects a different score per 𝜌 ; both component scores’ 𝜌𝑠 are reported above.
oracle gain, peak end-to-end [email protected] of 0.808, and AUC𝜌 = 0.795 over the offload-budget sweep. It outperforms the strongest weakconditioned baseline we could construct (XGBoost on MORIC: 0.472 / 0.804 / 0.794) on every metric, despite running before the weak detector and never seeing its features. Both EdgeML [23] and DCSB [3] sit well below either family. The MORIC+ regression target lands between the two families and gives a useful soft signal for the smaller-trunk variant. These results indicate that raw images are sufficient for objectdetection routing. With 0.15 GFLOPs of estimator compute, we obtain a routing signal that exceeds a strong proposal-feature baseline whose signal depends on the weak model’s 4.49 GFLOPs. This confirms the suspicion raised in § 1: predicting whether a frame is difficult is substantially cheaper than predicting what it contains.
4.2
The Compute Advantage is Material
Fig. 4 decomposes the expected per-frame cost for weak-conditioned (C) and weak-skipping (S) at each budget 𝜌 into three stacked components: estimator 𝐶𝑒 (blue), weak detector (orange), and cloud 𝜌𝐶𝑠 (gray). The cloud band dominates both placements equally (𝜌𝐶𝑠 ), so the saving comes entirely from the device side: in the S bars the orange weak-detector band shrinks with (1−𝜌)𝐶 𝑤 instead of the constant 𝐶 𝑤 paid in C, at the cost of adding a thin blue estimator band (𝐶𝑒 =0.15 GFLOPs). Compute (panel a). The GFLOP saving is monotone in 𝜌: from 0.3 GFLOPs at 𝜌=0.1 through 1.6 at 𝜌=0.4 and 2.5 at 𝜌=0.6, to 3.9 at 𝜌=0.9, approaching one full weak pass (4.49 GFLOPs) as 𝜌→1. The breakeven is very low, 𝜌 ∗ =𝐶𝑒 /𝐶 𝑤 ≈0.03: weak-skipping is cheaper for essentially all operating budgets, since the estimator pays for itself once it avoids more than one weak pass in ∼30 frames.
Budget-Adaptive Routing: Skipping the Weak When the Strong Answers Anyway
SIGCOMM ’26, August 17–21, 2026, Denver, CO, USA
(a) Compute (GFLOPs / frame)
GFLOPs / frame
200
Cloud ρCs Weak detector Estimator Ce
100 −0.3
−0.7
−1.2
−1.6
−2.1
−2.5
−3.0
−3.4
(b) Latency (ms / frame) −3.9
Cloud ρCs Weak detector Estimator Ce
70 60
ms / frame
300
10
50 40 30
+0.6
−4.3
−1.9
−6.8
−9.3
−11.7
−14.2
−16.7
−19.1
20 5 0
10 C S
C S
C S
C S
C S
C S
C S
C S
C S
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
C S
C S
C S
C S
C S
C S
C S
C S
C S
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Offload budget ρ (C=Cond, S=Skip)
Offload budget ρ (C=Cond, S=Skip)
Figure 4: Expected per-frame cost vs. offload budget 𝜌 on VOC. Each bar stacks estimator 𝐶𝑒 (blue), weak detector (orange), and cloud 𝜌𝐶𝑠 (gray). Paired bars are C (weak-conditioned, weak on every frame) and S (weak-skipping, weak on only the (1−𝜌) kept frames); each pair’s label is the cond−skip saving (pink when weak-skipping costs more). (a) Compute (GFLOPs, broken 𝑦-axis; 𝐶𝑒 too small to see). (b) Latency (ms), which adds the serial weak-pass dependency.
4.3
Budget-adaptive Achieves Upper Envelope
(1) Neither fixed placement dominates the operating curve. Fig. 5 shows weak-conditioned (XGBoost on MORIC) mAP leads at low and very high budgets (𝜌≤0.2 and 𝜌≥0.8, by 0.1–0.4 pp), while weakskipping (MV2-Lite + OffloadBin) dominates in the mid-budget regime (𝜌∈{0.3, . . . , 0.7}). Tab. 3 shows that two crossovers bracket the skipping band: at 𝜌 frontier ≈0.3 the weak forward pass on offloaded frames becomes a pure tax and skipping it yields both compute savings and better frame selection; at 𝜌 ceiling ≈0.8 the few remaining local frames carry enough weight that the richer weakconditioned signal again dominates the per-frame compute saved by skipping, also shown in Fig. 5 frame selection quality-wise. (2) Budget-adaptive traces the upper envelope of both placements and strictly dominates all baselines. A two-threshold offline arbiter (𝜌 frontier =0.3, 𝜌 ceiling =0.8) produces the per-budget maximum (last row of Tab. 3, teal curve in Fig. 5): it picks weak-conditioned at 𝜌≤0.2 and 𝜌≥0.8, weak-skipping at 0.3≤𝜌≤0.7. This envelope achieves the highest peak [email protected] of any router, 0.808, which is +1.7 pp over the strong model itself, and the highest AUC𝜌 =0.796 (Tab. 2). Compared with prior weak-conditioned SOTAs, the budget-adaptive system leads EdgeML [23] by 0.7–2.1 pp across the full budget range (the strong model itself leads the weak model by only 3.1 pp). The
0.83 0.82
End-to-end [email protected]
Latency (panel b). The wall-clock picture differs because it captures a serial dependency. A weak-conditioned router cannot produce a score until the weak forward pass finishes (24.70 ms), whereas a weak-skipping router runs 𝑓skip (3.08 ms) in its place on offloaded frames. At 𝜌=0.1 weak-skipping is 0.6 ms slower (pink label) because the estimator overhead slightly exceeds the small saving from skipping only 10% of weak passes. The crossover sits ∗ ≈0.12. Beyond it the saving grows near-linearly, up to ∼30% at 𝜌 ms end-to-end latency reduction. For latency-critical deployments with higher 𝜌 , this saving alone justifies the weak-skipping placement.
0.81 0.80 0.79
Strong only Oracle Random EdgeML (cond) DCSB (cond, fixed) XGBoost+MORIC (cond, ours) MV2+MORIC+ (skip, ours) MV2+OffloadBin (skip, ours) Adaptive (envelope, ours)
0.78 0.77 0.76
Weak only
ρfrontier ≈ 0.30
0.0
0.2
0.4
0.6
0.8
1.0
Offload budget ρ
Figure 5: End-to-end [email protected] vs. offload budget 𝜌 on VOC. budget-adaptive (teal) is the pointwise upper hull of weak-skipping (blue) and weak-conditioned (orange), with crossovers at 𝜌 frontier ≈0.3, 𝜌 ceiling ≈0.8.
widest gap is at 𝜌=0.2 (0.786 vs. 0.765) and the narrowest is at 𝜌=0.9 (0.798 vs. 0.791) where EdgeML’s native threshold saturates to always-offload. DCSB [3] is a fixed binary rule locked to a single operating point (𝜌≈0.79, mAP= 0.789). At the same budget the budget-adaptive router reaches 0.803, and its peak (0.808) exceeds DCSB by 1.9 pp while offering continuous budget tunability. The envelope sits 1.9 ∼ 2.9 pp below the offline oracle, bounding what routing quality alone can recover.
SIGCOMM ’26, August 17–21, 2026, Denver, CO, USA
Geng et al.
Table 3: End-to-end [email protected] on VOC across offload budgets 𝜌 ∈ {0.1, . . . , 0.9} (single seed). Each cell is the accuracy at that 𝜌. ▲ beats and • matches the always-strong model (0.791), an unmarked cell is below it. Bold is the column best. On the budget-adaptive row a superscript marks the selected branch (S = weak-skipping, C = weak-conditioned). 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Reference Weak only (constant) Strong only (constant) Random offloader Oracle (ΔAP)
0.760 0.791 0.765 0.799 ▲
0.760 0.791 0.769 0.815 ▲
0.760 0.791 0.772 0.823 ▲
0.760 0.791 0.776 0.827 ▲
0.760 0.791 0.778 0.827 ▲
0.760 0.791 0.781 0.827 ▲
0.760 0.791 0.784 0.826 ▲
0.760 0.791 0.786 0.824 ▲
0.760 0.791 0.788 0.817 ▲
Weak-conditioned EdgeML [23] XGBoost on MORIC (Ours)
0.760 0.776
0.765 0.786
0.789 0.794 ▲
0.793 ▲ 0.799 ▲
0.791 • 0.803 ▲
0.791 • 0.804 ▲
0.791 • 0.803 ▲
0.791 • 0.803 ▲
0.791 • 0.798 ▲
Weak-skipping MV2 + MORIC+ (Ours) MV2 + OffloadBin (Ours)
0.771 0.773
0.778 0.785
0.785 0.799 ▲
0.791 • 0.806 ▲
0.796 ▲ 0.808 ▲
0.800 ▲ 0.805 ▲
0.801 ▲ 0.804 ▲
0.801 ▲ 0.799 ▲
0.798 ▲ 0.794 ▲
Budget-adaptive (regime in superscript) MV2/OffloadBin ↔ XGB/MORIC (Ours)
0.776C
0.786C
0.799S ▲
0.806S ▲
0.808S ▲ 0.805S ▲
0.804S ▲ 0.803C ▲
0.798C ▲
𝜌
(3) The dominance extends to the joint accuracy-compute frontier inside the skipping band. For 0.3≤𝜌≤0.7, budget-adaptive inherits weak-skipping’s compute profile: 2.1 GFLOPs/frame less than any weak-conditioned method at 𝜌=0.5 and 3.0 GFLOPs/frame less at 𝜌=0.7 (the top of the band; Fig. 4a), with wall-clock savings reaching 14.2 ms (∼26% reduction) at 𝜌=0.7 (Fig. 4b). At 𝜌<𝜌 frontier the weak pass runs on nearly all frames regardless, so switching to weakconditioned costs no incremental compute. At 𝜌≥𝜌 ceiling budgetadaptive has the option to trade the skipping compute saving for the 0.4 pp accuracy lift of weak-conditioned. Inside the skipping band the budget-adaptive router is therefore never worse in accuracy and never worse in compute than either fixed placement alone.
5
Related Work
Selective-offloading admits a clean taxonomy along one architectural axis: when does the estimator fire, relative to the weak model? Weak-conditioned estimators. Shown in Fig. 2-middle, the estimator fires after the weak model and consumes its outputs. EdgeML [23] regresses on the top-25 proposal features and DCSB [3] hand-crafts a difficult-case discriminator. Confidence thresholds and learned detection embeddings [28] also belong to this class. They exploit rich signals but pay implicit compute tax unconditionally. Mid-network gates and early exit. This is essentially weak-conditioned estimators. BranchyNet [26], MSDNet [12], and SDN [15] fuse the estimator into the weak network and gate at intermediate layers. They reduce average weak-model cost on easy frames but do not skip the "Early Pass" entirely. Neurosurgeon [14] and SPINN [16, 17, 22] partition or progressively split a single network across device and cloud, which is related but distinct. Cascaded inference. Classical cascades [27] and modern cascade detectors [1] escalate on uncertainty. As detailed in § 3.2, they use the cheap stage to produce predictions whereas budget-adaptive routing uses it to route.
Weak-skipping estimators. The estimator fires before the weak model from the raw image. Early image-complexity heuristics and learned image-level routers in our work belong to this class. They allow the weak pass to be skipped entirely on offloaded frames but have often been considered too weak for detection routing, a premise challenged by our results.
6
Conclusion and Future Work
Estimator placement is a design axis the selective-offloading literature has implicitly fixed. We exhibited both a lightweight imageonly weak-skipping estimator and a weak-conditioned estimator that outperform the SOTAs that we could reproduce from the literature, on both routing quality and compute. We showed the placement choice is regime-dependent and built a budget-aware budgetadaptive router that traces the upper accuracy envelope of both placements on VOC while saving compute and latency within its skipping band. Several directions remain. A real edge-cloud testbed would replace our offline simulation with measured round-trips [7] on real-world hardwares. Testing beyond VOC such as COCO[21] would probe how far the raw-pixel routing signal carries. Beyond [email protected], stricter AP@[.5 : .95], class-weighted or downstreamtask utility will be experimented. A skip-then-reconsider cascade would run the weak-skipping estimator first and, on frames kept local where the weak pass executes anyway, apply a weak-conditioned estimator to revisit the offload decision on the now-available weak features at no extra detector cost. A head-to-head with early-exit routers across weak-model prefix depths would chart when a shallow prefix already rivals the image-only weak-skipping estimator.
Acknowledgments This work was supported by the Dutch National Growth Fund “Future Network Services”.
Budget-Adaptive Routing: Skipping the Weak When the Strong Answers Anyway
References [1] Zhaowei Cai and Nuno Vasconcelos. 2018. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition. 6154–6162. [2] Christopher Canel, Thomas Kim, Giulio Zhou, Conglong Li, Hyeontaek Lim, David G Andersen, Michael Kaminsky, and Subramanya Dulloor. 2019. Scaling Video Analytics on Constrained Edge Nodes. In Proceedings of Machine Learning and Systems, A. Talwalkar, V. Smith, and M. Zaharia (Eds.), Vol. 1. 406–417. https://proceedings.mlsys.org/paper_files/paper/2019/file/ 6bcfac823d40046dca25ef6d6d59cc3f-Paper.pdf [3] Zhiqiang Cao, Zhijun Li, Yongrui Chen, Heng Pan, Youbing Hu, and Jie Liu. 2023. Edge-cloud collaborated object detection via difficult-case discriminator. In 2023 IEEE 43rd International Conference on Distributed Computing Systems (ICDCS). IEEE, 259–270. [4] Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 785–794. [5] Tiffany Yu-Han Chen, Lenin Ravindranath, Shuo Deng, Paramvir Bahl, and Hari Balakrishnan. 2015. Glimpse: Continuous, real-time object recognition on mobile devices. In Proceedings of the 13th ACM conference on embedded networked sensor systems. 155–168. [6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. 2010. The pascal visual object classes (voc) challenge. International journal of computer vision 88, 2 (2010), 303–338. [7] Wei Geng, Oguz Kagan Altas, David Guzman, Giovanni Bartolomeo, Nitinder Mohan, and Joerg Ott. 2025. Poster: KUT: Towards Lightweight On-path Network Assessment for Edge Orchestration. In Proceedings of the 21st International Conference on emerging Networking EXperiments and Technologies. 9–11. [8] Wei Geng, Xiang Su, Nitinder Mohan, Jörg Ott, and Pan Hui. 2026. SMOOTH: Scalable Multitask Offloading with Backbone Sharing. In 2026 IFIP Networking Conference (IFIP Networking). IFIP, 1–10. [9] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning. PMLR, 1321–1330. [10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition. 770–778. [11] Andrew Howard, Mark Sandler, Bo Chen, et al. 2019. Searching for MobileNetV3. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). [12] Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens Van Der Maaten, and Kilian Q Weinberger. 2017. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844 (2017). [13] Junchen Jiang, Ganesh Ananthanarayanan, Peter Bodik, Siddhartha Sen, and Ion Stoica. 2018. Chameleon: scalable adaptation of video analytics. In Proceedings of the 2018 conference of the ACM special interest group on data communication. 253–266. [14] Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang. 2017. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. ACM SIGARCH Computer Architecture News 45, 1 (2017), 615–629. [15] Yigitcan Kaya, Sanghyun Hong, and Tudor Dumitras. 2019. Shallow-deep networks: Understanding and mitigating network overthinking. In International conference on machine learning. PMLR, 3301–3310. [16] Stefanos Laskaridis, Stylianos I Venieris, Mario Almeida, Ilias Leontiadis, and Nicholas D Lane. 2020. SPINN: Synergistic progressive inference of neural networks over device and cloud. In Proceedings of the 26th annual international conference on mobile computing and networking. 1–15. [17] En Li, Zhi Zhou, and Xu Chen. 2018. Edge intelligence: On-demand deep learning model co-inference with device-edge synergy. In Proceedings of the 2018 workshop on mobile edge communications. 31–36. [18] Yuanqi Li, Arthi Padmanabhan, Pengzhan Zhao, Yufei Wang, Guoqing Harry Xu, and Ravi Netravali. 2020. Reducto: On-camera filtering for resource-efficient real-time video analytics. In Proceedings of the Annual conference of the ACM Special Interest Group on Data Communication on the applications, technologies, architectures, and protocols for computer communication. 359–376. [19] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. 2017. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition. 2117–2125. [20] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision. 2980–2988. [21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. 2014. Microsoft COCO: Common Objects in Context. In Proceedings of the European Conference on Computer Vision (ECCV).
SIGCOMM ’26, August 17–21, 2026, Denver, CO, USA
[22] Yoshitomo Matsubara, Marco Levorato, and Francesco Restuccia. 2022. Split computing and early exiting for deep learning applications: Survey and research challenges. Comput. Surveys 55, 5 (2022), 1–30. [23] Jiaming Qiu, Ruiqi Wang, Brooks Hu, Roch Guérin, and Chenyang Lu. 2024. Optimizing edge offloading decisions for object detection. In 2024 IEEE/ACM Symposium on Edge Computing (SEC). IEEE, 164–177. [24] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. 2015. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In Advances in Neural Information Processing Systems (NeurIPS). [25] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and LiangChieh Chen. 2018. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition. 4510–4520. [26] Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. 2016. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd international conference on pattern recognition (ICPR). IEEE, 2464–2469. [27] Paul Viola and Michael Jones. 2001. Rapid object detection using a boosted cascade of simple features. In Proceedings of the 2001 IEEE computer society conference on computer vision and pattern recognition. CVPR 2001, Vol. 1. Ieee, I–I. [28] Qingyuan Wang, Barry Cardiff, Antoine Frappé, Benoit Larras, and Deepu John. 2024. Tiny models are the computational saver for large models. In European Conference on Computer Vision. Springer, 163–182.
A
Supplementary Evidence
All numbers below are from the same single-pass VOC evaluation (seed 42) used in § 4, profiled on the setup of Tab. 1. Peak mAP is the maximum end-to-end accuracy over the offload-budget 𝜌 sweep. [email protected] is the VOC metric and AP@[.5:.95] (COCO-style; written APC in table headers) is the stricter metric. These appendices backs and justifies three design choices made in § 3 (the learning target, the estimator backbone, and the claim that routing signal is recoverable from the raw image) and probe robustness to a stricter accuracy metric.
A.1
Backbone Ablation
Table 4: Backbone ablation for the weak-skipping estimator, controlling for the learning target: both trunks use the identical OffloadBin/focal target and differ only in the backbone (VOC test, seed 42). GFLOPs and parameters are for the estimator alone. Trunks we trained on other targets are not directly comparable and are omitted. Backbone (OffloadBin/focal)
GFLOPs
Par. (M)
𝜌𝑠
Peak mAP
MobileNetV2-Lite EfficientNet-B0-Lite
0.153 0.176
0.54 0.85
0.557 0.235
0.808 0.795
Our backbone sweep is small, so we report it only briefly. Tab. 4 holds the learning target fixed at OffloadBin/focal and varies the trunk alone: the larger EfficientNet-B0-Lite does not improve routing over the compact MobileNetV2-Lite, with a markedly lower rank correlation and no gain in peak mAP. This is consistent with the premise of § 1 that detecting difficulty needs little capacity, though a broader sweep is left to future work.
A.2
Learning-Target Ablation
§ 3.1 reduces the long-tailed per-frame ΔAP to a binary OffloadBin label trained with focal loss, rather than regressing a continuous reward. Because the backbone and the target must be chosen jointly, a chicken-and-egg dependency, Tab. 5 fixes the backbone first and sweeps the target. OffloadBin wins by a wide margin (𝜌𝑠 =0.557,
SIGCOMM ’26, August 17–21, 2026, Denver, CO, USA
Geng et al.
Table 5: Learning-target ablation on the fixed MobileNetV2-Lite weak-skipping backbone (VOC test, seed 42). 𝜌𝑠 is the Spearman correlation between the estimator score and the per-frame oracle gain (not the offload budget 𝜌), and ratio-err is the mean absolute gap between realized and requested offload fractions. Exact target definitions are in our released artifacts. Rows are sorted by 𝜌𝑠 within each group, best per column in bold. Target / loss
𝜌𝑠
Peak mAP
Peak APC
ratio-err
Classification target (focal) OffloadBin Binary 1[ΔAP>0] : does the cloud strictly help this frame? Focal loss handles the positive-class imbalance. TopQuartile Binary: is the frame’s gain in the top quartile of ΔAP? A rarer, harder positive class than OffloadBin.
Description
0.557 0.389
0.808 0.800
0.609 0.606
0.009 0.007
Regression target MORIC+ HighIoUGain MORIC+ (quantile) F1Gain MORIC+ (wing) RescueRatio WorstCaseGain SigMORIC MORIC★ Φ-MORIC RescueRatio (wing)
0.350 0.345 0.313 0.306 0.303 0.291 0.281 0.273 0.271 0.263 0.169
0.801 0.794 0.801 0.791 0.799 0.792 0.796 0.797 0.797 0.797 0.791
0.607 0.607 0.608 0.605 0.606 0.605 0.607 0.605 0.605 0.605 0.605
0.018 0.008 0.010 0.006 0.024 0.004 0.010 0.015 0.015 0.015 0.011
Signed empirical CDF of ΔAP, split at zero onto [−1, 1] ; this is our reward (Eq. (5)). Continuous per-frame gain proxy that up-weights tightly-localised, high-IoU matches. The MORIC+ target, fit with a quantile (pinball) regression loss. Continuous per-frame proxy: the change in detection F1 when the frame is offloaded. The MORIC+ target, fit with a wing regression loss. Continuous proxy for the share of a frame’s missed objects that the cloud recovers. Continuous proxy that emphasises the frame’s worst-case detection loss. The MORIC+ reward passed through a sigmoid squashing. The MORIC+ reward under an alternative CDF reshaping. The MORIC+ reward under a Φ-based CDF reshaping. The RescueRatio target, fit with a wing regression loss.
low → high confidence
(a) Weak confidence
mAP gain (pp)
39%
Q1 Q2
31%
Q3
25%
Q4
6%
7.1 3.7 2.5 1.2
0
25
50
75
100%
(b) Scene crowding sparse → crowded
Q1
9%
Q2
22%
2.7 4.1
34%
Q3
4.6
47%
Q4 0
mAP gain (pp)
25 Benefit
4.0
50 75 100% Harm Neutral
Figure 6: Where offloading helps, by weak-detector stratum (VOC test, seed 42, 𝑁 =3105). Each bar splits a quartile’s frames into Benefit (offloading raises ΔAP), Harm (lowers it), and Neutral (unchanged). The dominant neutral mass shows the benefit is a sparse partition. (a) Across weak-confidence quartiles benefit falls 6.4× (Q1→Q4). (b) Across scene crowding (weak detection count) it rises 5.1×. The right strip is the perframe weak→strong mAP gain in points (pp), largest where benefit concentrates. against 0.389 for the next-best target and 0.169 ∼ 0.350 for the continuous-regression variants) and tops both accuracy metrics. Budget tracking is uniformly tight, so this gap reflects the target itself rather than calibration. The result confirms the § 3.1 argument: collapsing whether offloading helps into a binary label decouples
hardness from content and lets raw pixels suffice, whereas regressing the signed magnitude wastes capacity on the 62.6% zero spike (Fig. 3).
A.3
Where Offloading Helps
A weak-skipping estimator presumes that which frames benefit from the cloud is both structured and recoverable from the raw image. Fig. 6 speaks to the first half: the benefit is sharply concentrated, with the lowest weak-confidence quartile offload-beneficial 6.4× more often than the highest (38.9% vs. 6.1%) and the most crowded scenes 5.1× more often than the sparsest (46.9% vs. 9.2%), tabulated in full in Tab. 6. The same strata carry the largest weak→strong mAP headroom. These strata are defined by weak-detector outputs, however, so these views localise where offloading helps rather than showing the signal is recoverable before the weak pass. We assume that low confidence and crowding are correlates of scene complexity Table 6: Offload benefit broken down by weak-detector stratum on the VOC test set (seed 42, 𝑁 =3105 frames). We split the frames into quartiles Q1–Q4 of each weak-detector summary, so that Q1 contains the least-confident or sparsest scenes and Q4 the most-confident or most-crowded. The Benefit, Harm, and Neutral columns give the fraction of frames in each stratum for which offloading respectively raises, lowers, or leaves ΔAP unchanged, and they sum to one. The final two columns, mAP𝑤 and mAP𝑠 , report the mean per-frame mAP of the weak and strong detectors within the stratum. Stratum
Q
Benefit
Harm
Neutral
mAP𝑤
mAP𝑠
Weak conf. (mean)
Q1 Q2 Q3 Q4
0.389 0.312 0.246 0.061
0.169 0.173 0.121 0.030
0.443 0.515 0.633 0.911
0.750 0.847 0.895 0.961
0.821 0.884 0.920 0.973
Weak det. count
Q1 Q2 Q3 Q4
0.092 0.217 0.343 0.469
0.037 0.093 0.151 0.269
0.872 0.690 0.507 0.262
0.928 0.864 0.831 0.784
0.955 0.905 0.877 0.824
Budget-Adaptive Routing: Skipping the Weak When the Strong Answers Anyway
that is plausibly legible from raw pixels and is established directly by the weak-skipping estimator’s measured routing quality (Tab. 5). The benefit structure dictates the target. Tab. 5 and Fig. 6 are two views of one phenomenon. The offload benefit is a sparse partition, not a smooth magnitude field: a dominant neutral mass (62.6% of frames at ΔAP=0; median gain 0 in every stratum) around a minority of beneficial frames (≤47% even at best) in low-confidence,
SIGCOMM ’26, August 17–21, 2026, Denver, CO, USA
crowded scenes. The learnable question is thus whether a frame lies in that region, not by how much it gains, which is why Tab. 5 ranks both classification targets (OffloadBin 𝜌𝑠 =0.557, TopQuartile 0.389) above every regression variant (𝜌𝑠 ≤0.350), and why routing reduces to recognising a region of image space that a cheap raw-image trunk can read.