Janus: Collaborative Vision Transformer Under Dynamic Network Environment

Bo Li; Linyi Jiang; Silvery D. Fu; Yifei Zhu

REVIEW 3 major objections 7 minor 47 references

Janus: Collaborative Vision Transformer Under Dynamic Network Environment

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Janus runs ViTs jointly on edge and cloud, choosing on the fly how many tokens to prune and where to split so latency stays within its target as the network fluctuates.

desk verdict A coherent and genuinely useful integration of token pruning with model splitting for collaborative ViT inference, but the headline performance numbers are likely computed by the scheduler's own profiler and need an independent end-to-end check before they can be trusted. read the letter →

arxiv 2502.10047 v1 pith:LWMFQMA3 submitted 2025-02-14 cs.DC cs.AI

Linyi Jiang , Silvery D. Fu , Yifei Zhu , Bo Li This is my paper · ORCID

classification cs.DCcs.AI

keywords VisionTransformercloud-devicecollaborationmodelsplittingtokenpruningdynamicnetworkslatencyoptimizationedgeinferencevideoanalytics

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Vision Transformers are accurate but expensive, and a plain ViT sends tensors of unchanged size across the network, so neither running it entirely on a phone nor entirely in the cloud fits when bandwidth swings. Janus claims the first low-latency collaborative scheme for ViTs: token pruning shrinks the intermediate data before transmission, model splitting partitions the rest of the computation between device and cloud, and a lightweight scheduler picks the pruning level and split point that meet a latency deadline under the current network. Experiments on ImageNet-1k image classification and Kinetics-400 video classification, driven by real 4G/5G uplink traces and a live WiFi/4G/5G deployment, report up to 5.15x higher throughput and up to 98.7% fewer latency-requirement violations than device-only, cloud-only, and mixed baselines, with average accuracy within 0.29% of those baselines. The point of the system is to make high-accuracy ViT inference viable on resource-limited edge devices even when network quality varies.

What carries the argument

The load-bearing object is the mixed pruning policy with an exponential declining rate $\alpha$: for layer $l$ of an $N$-layer ViT, the number of pruned tokens is $\Delta x_l = \lfloor 2\alpha(N-l)\rfloor$ when $\alpha>0$, and zero when $\alpha=0$, so early layers lose many tokens and later layers lose few. This gives the scheduler a one-dimensional accuracy-versus-latency knob that feeds the profiler's linear per-layer latency models $f^M_{\text{device}}(x_l)$ and $f^M_{\text{cloud}}(x_l)$, which predict latency from the token count $x_l$. The scheduler minimizes $L_{s,\alpha}=\sum_{l=1}^{s}T^{\text{device}}_l+\sum_{l=s+1}^{N}T^{\text{cloud}}_l+x_sD_M/B$ over the candidate split points produced by a fine-to-coarse splitter, which places more candidates in the front layers (where pruning has already delivered the largest communication saving) and always includes the cloud-only and device-only endpoints. This design keeps the search small enough that the paper reports an average scheduler execution time of about 1 ms.

What would settle it

Measure per-layer latency of ViT-L on the edge device used in the paper for token counts from 1 to the full sequence, fit the profiler's linear model, and check whether residuals grow at small token counts; if they do, the scheduler's predicted latency for aggressive pruning is systematically wrong.

Watch

Extended reading notes

Core claim

The paper's central claim is that ViTs can be made collaborative after all, once token pruning is used to create the data reduction that vanilla transformer layers lack. In a ViT the intermediate tensor stays the same size at every layer, which is why prior model-splitting schemes work for CNNs but not for ViTs; Janus removes tokens early in the network so that the tensor crossing to the cloud is genuinely smaller. The scheduler then chooses how aggressive the pruning should be and where the model should be cut, scanning from no pruning to the maximum allowed declining rate and returning the highest-accuracy configuration whose predicted end-to-end latency satisfies the service-level agreement. In the reported experiments, the resulting behavior is to offload to the cloud without pruning when bandwidth is good, and to prune and split as bandwidth degrades, which is what keeps average throughput high and latency violation ratios near zero.

Load-bearing premise

The scheduler's value rests on two estimates being accurate enough: that each layer's latency grows linearly with its input token count, and that the recent average bandwidth estimate tracks the current network speed; if either is wrong, Janus can choose a pruning level or split point that misses the latency target.

Editorial extensions

If this is right

When the network is good, Janus's scheduler selects cloud-only execution with no pruning, so on strong links the system behaves like a plain cloud service and preserves full accuracy.
When bandwidth drops, the same scheduler increases the declining rate and moves the split toward the device, trading a controlled amount of accuracy to keep the latency deadline.
For video classification, the same mechanism applies at clip level, so Janus extends from frames to spatiotemporal ViTs without changing the core scheduling logic.
Because the scheduler runs in about 1 ms and the paper reports total system overhead below 0.21% of end-to-end latency, the planning cost does not eat the latency savings.
The reported accuracy gap to baselines is at most a few tenths of a percent under dynamic token budgets, which implies that adaptive pruning can buy most of the latency gain without a visible accuracy penalty.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next test is sensitivity to profiler error: if layer latency departs from the linear model at very small token counts, the scheduler will overestimate the benefit of aggressive pruning, and the paper does not include that sensitivity analysis.
The same split-and-prune logic may transfer to other transformer workloads, such as language models, where token pruning before a split point could shrink the context that must cross the network.
The measured accuracy advantage over fixed-pruning baselines likely reflects that adaptive pruning keeps more tokens when bandwidth is plentiful and only prunes hard when the deadline demands it.
In very bursty networks, the harmonic-mean bandwidth estimate may lag the instantaneous rate, so a scheduler variant that probes current bandwidth before each decision could reveal whether remaining violations come from estimation lag rather than pruning policy.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First, what's new: Janus combines token pruning with model splitting for ViTs, with a scheduler that picks both pruning intensity and split point based on a lightweight linear profiler and current bandwidth. The front-loaded (exponential) declining pruning profile is a sensible design choice, and the fine-to-coarse split candidate set is a reasonable way to cut scheduler overhead. The paper is honest about its components being known; the integration is the contribution.

What it does well: the system design is coherent, the motivation is clearly measured (the latency breakdown in Fig. 2 is useful), and the real prototype with a Jetson and cloud GPU is real. The overhead analysis (Table II) is a legitimate strength — scheduler overhead under 0.21% is a meaningful engineering result. The paper also cites prior work adequately; I don't see obvious citation gaming.

The soft spots are real but specific. The headline numbers (5.15x throughput, 98.7% violation-ratio reduction) come from trace-driven simulation. The natural reading of Section V-C is that the simulation uses the same linear profiler both to choose the configuration and to compute the achieved latency. If that's true, the scheduler is graded against its own predictions, and any systematic profiler error is invisible in the reported metrics. The paper never states that achieved latency in simulation is measured from real device/cloud execution. That's the load-bearing issue, and it sharpens the reader's profiler-sensitivity concern: the paper also doesn't test sensitivity to profiler error or bandwidth estimation error. The accuracy claim (within 0.29%) is only relative to baselines, not absolute; that's a less serious gap but worth noting. No code or data released, which in this subfield is common but makes the simulation hard to check.

I think the stress-test note is right on this. I'd rank the circularity concern as the main reason not to accept the numbers at face value. That said, the core idea — that token pruning creates the data reduction that makes splitting worthwhile for ViTs — is sound and likely to be useful regardless of the exact magnitude of the gains. The paper deserves a serious referee, not a desk reject. It's a solid systems paper with a fixable evaluation gap.

Who's this for: systems folks working on edge-cloud inference, especially video analytics. I'd bring it to a reading group discussion about evaluation methodology in systems papers, but I wouldn't cite the headline numbers without independent validation. Recommendation: send to peer review with a request for end-to-end measured latency in the trace simulation, or at least a clear statement of how simulation latency is computed, plus a profiler-error sensitivity analysis.

Referee Report

3 major / 7 minor

Summary. Janus is a cloud-device collaborative inference framework for Vision Transformers (ViTs). It combines token pruning with model splitting: a collaboration-aware token pruner removes tokens using a "declining rate" policy, a fine-to-coarse splitter restricts candidate split points, a lightweight linear profiler predicts per-layer latency as a function of token count, and a dynamic scheduler selects the declining rate and split point to meet a latency SLA under an estimated bandwidth. The paper reports trace-driven simulation results on 4G/5G traces for image recognition (ImageNet with ViT-L@384) and video classification (Kinetics-400 with Spatiotemporal MAE ViT-L), claiming up to 5.15x throughput improvement and up to 98.7% reduction in latency violation ratio over Device-Only, Cloud-Only, and Mixed baselines, with accuracy within 0.29% of baselines. A real prototype is implemented on Jetson Orin Nano and an Aliyun V100 cloud instance, with an overhead analysis showing system overhead below 0.21%.

Significance. If the reported gains are robust, the work addresses a genuine and timely problem: unlike CNNs, vanilla ViTs do not reduce intermediate data size, so existing model-splitting approaches are ineffective. The idea of using token pruning to create data reduction at split points and dynamically choosing split and pruning levels is novel and practically motivated. The paper ships a real prototype and a low-overhead scheduler, and the design is well integrated. However, the central quantitative claims rest on a trace-driven simulation that may use the same latency model the scheduler itself relies on, which is a load-bearing circularity risk. The accuracy evaluation is also reported only relative to baselines. These issues need to be resolved before the headline numbers can be taken as established.

major comments (3)

[Section V-C, Algorithm 1] The throughput and violation-ratio results are obtained from a trace-driven simulation, but the manuscript does not state how the achieved latency is computed in that simulation. If the simulation scores each configuration using the same linear profiler f_M and harmonic-mean bandwidth estimate that Algorithm 1 uses to select the declining rate and split point, then the scheduler is being evaluated against its own predictions; a systematically wrong profiler would then affect the decision and the reported latency in the same way, masking the error. Please clarify the simulation's latency model, and ideally report end-to-end measured latency from the real prototype on a replay of the same traces, or at least quantify profiler prediction error against actual layer execution for the configurations chosen and show that the reported violation ratios and throughput gains are robust to that error.
[Section III-D, Algorithm 1 line 9] The bandwidth estimate is the harmonic mean of observed throughput, which can lag during sudden drops in the dynamic traces used in Section V-C (e.g., LTE Driving). The sensitivity analysis in Section V-D varies only constant bandwidth values and does not exercise the estimator's error or lag. Since the scheduler's split and pruning decisions depend directly on this estimate, the paper should include a sensitivity analysis with delayed or biased bandwidth estimates, or a real deployment under a dynamic network, to demonstrate that the reported improvements are not an artifact of a well-conditioned bandwidth estimator.
[Section V-B/V-C, Fig. 7] Accuracy is reported only as an average improvement of 0.01% to 0.29% over the baselines, which does not allow the reader to see the absolute accuracy cost of the latency-reduction mechanism. Please report absolute top-1 accuracy for Janus and each baseline for both tasks, and ideally an accuracy-latency tradeoff curve over the range of declining rates α, to substantiate the claim of "minimal accuracy reduction" relative to the unpruned model and to make the tradeoff explicit.

minor comments (7)

[Section III-A, Eq. (1)] The pruning policy is called "exponential-form" but the formula ∆xl = floor(2α(N-l)) is linear in (N-l) and differs from the "linear declining" comparison only by a factor of 2. Please revise the terminology or the formula so that the description matches the actual functional form.
[Section III-A, Eq. (1) vs Eq. (2)] Equation (1) defines pruned tokens at layer l as floor(2α(N-l)), while Equation (2) uses N-(l-1) = N-l+1 in the sum constraint. This off-by-one inconsistency should be aligned so the cumulative pruning count is computed from the same formula.
[Section III-B, Eq. (3)] The recurrence si = si-1 + floor(i/k) yields duplicate values when floor(i/k)=0 (e.g., for i<k). While the set notation may collapse duplicates, the paper should clarify that the sequence itself is not strictly increasing and define the resulting unique candidate set explicitly.
[Section V-C, Fig. 7] The ranges for throughput improvement and violation-ratio reduction are wide (e.g., 1.00x to 5.15x and 49.8% to 98.7%); please specify which baseline, network type, and task corresponds to each extreme so the reader can interpret the headline numbers precisely.
[Section V-D, Fig. 9] The latency curves in Fig. 9 appear to be predicted values from the profiler; please state explicitly whether these are measured or predicted latencies, and clarify how the profiler-based predictions compare with actual execution.
[General] The paper claims to be the "first" collaborative ViT inference framework in the abstract, Section I, and Section VI; please verify this claim against concurrent work on split inference for transformers and soften the wording if any prior work exists.
[Figure 3] There is a typo in Figure 3: "Model Spliter" should be "Model Splitter."

Circularity Check

1 steps flagged · score 6.0 of 10

Headline throughput and violation-ratio gains are scored by the same fitted profiler that makes the scheduling decisions, so the central simulation results are not independently validated.

fitted input called prediction [Section III-C/III-D (Algorithm 1) and Section V-C simulation]
"Using prediction models in the profiler, the algorithm predicts the device latency T device l for xl tokens running on the device and the cloud latency T cloud l for the same layer running on the server (lines 6-7). ... We answer the first question by performing a simulation experiment on real-world dynamic network trace on different ViTs inference tasks (§V-C)."

The profiler is itself fitted ('We use linear regression to calculate the coefficient for this model,' Sec. III-C). Algorithm 1 selects the declining rate and split point by minimizing T_device + T_cloud + T_comm using that fitted f_M and the harmonic-mean bandwidth estimate, and the Fig. 7 throughput and violation-ratio results come from the Sec. V-C trace simulation, which is not described as measuring real end-to-end device/cloud execution. The only compute-latency source in that loop is f_M, so the scheduler is scored by its own predictor: for any configuration the scheduler returns, the f_M-computed compute portion satisfies the SLA by construction, profiler error is invisible, and the reported gains are partly self-consistent predictions rather than independently measured outcomes.

full rationale

The central quantitative claims (up to 5.15x throughput, up to 98.7% violation-ratio reduction) are supported by the trace-driven simulation described in Section V-C. Algorithm 1 uses the linear profiler f_M, fitted by linear regression to measured per-layer latencies, plus a harmonic-mean bandwidth estimate, to choose the pruning level and split point. If, as the paper's text implies, the simulation also scores achieved latency with the same f_M, then the evaluation is partly circular: the scheduler is being measured against its own predictions, and any configuration it returns already has predicted compute latency no greater than the SLA at decision time. The real-deployment data (Table II) report overhead breakdown percentages, not end-to-end latency or violation ratios, so they do not independently validate the headline gains. This is partial circularity rather than total equivalence: the network traces are real, the profiler is calibrated to hardware measurements, accuracy is genuinely evaluated, and there is no load-bearing self-citation or uniqueness theorem imported from the authors' prior work. The paper would need either an end-to-end measured latency evaluation or a sensitivity analysis against profiler error to break the self-prediction loop.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

Janus leans on empirically fitted latency models, pre-trained token-pruning accuracy retention, bandwidth estimation, and hand-chosen hyperparameters. These are reasonable engineering assumptions but they are inputs to the scheduler, so the quantitative claims inherit their accuracy.

free parameters (4)

t (declining rate increment) = 0.01
Hand-set search granularity for alpha; not justified in the paper.
k (splitter density parameter) = 5
Controls candidate split point spacing; hand-set, no sensitivity analysis.
per-layer latency model slopes and intercepts = Not reported
Fitted by linear regression from measurements (Section III-C); used by scheduler to make decisions.
factor 2 in Eq. (1) = 2
Chosen to match the 'exponential' label; actual formula is linear in layer index.

assumptions (5)

domain assumption Layer inference latency is a linear function of the number of input tokens.
Empirically evaluated with correlation >0.85 (Section III-C), but treated as a general law for all pruning levels.
domain assumption Token pruning of a pre-trained ViT without fine-tuning preserves acceptable accuracy.
Relies on prior token pruning works (ToMe, ATS); Janus does not fine-tune after pruning.
domain assumption Bandwidth at decision time is well estimated by the harmonic mean of recent observed throughput.
Taken from FESTIVE; no sensitivity to estimation error is evaluated.
domain assumption The fine-to-coarse candidate split set contains a near-optimal split point.
Argued from latency-reduction distribution, not proven; may miss optimal splits in the rear of the model.
domain assumption Intermediate tensor sizes scale with token count after pruning, so communication latency is x_l * D_M / B.
Assumes LZW compression ratio is constant and overhead negligible; not measured per layer.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Janus: Collaborative Vision Transformer Under Dynamic Network Environment." pith.science (2026). https://pith.science/paper/LWMFQMA3

@misc{pith2026250210047,
  author       = {Pith},
  title        = {Pith review of: Janus: Collaborative Vision Transformer Under Dynamic Network Environment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LWMFQMA3}},
  note         = {Machine review of arXiv:2502.10047}
}

read the original abstract

Vision Transformers (ViTs) have outperformed traditional Convolutional Neural Network architectures and achieved state-of-the-art results in various computer vision tasks. Since ViTs are computationally expensive, the models either have to be pruned to run on resource-limited edge devices only or have to be executed on remote cloud servers after receiving the raw data transmitted over fluctuating networks. The resulting degraded performance or high latency all hinder their widespread applications. In this paper, we present Janus, the first framework for low-latency cloud-device collaborative Vision Transformer inference over dynamic networks. Janus overcomes the intrinsic model limitations of ViTs and realizes collaboratively executing ViT models on both cloud and edge devices, achieving low latency, high accuracy, and low communication overhead. Specifically, Janus judiciously combines token pruning techniques with a carefully designed fine-to-coarse model splitting policy and non-static mixed pruning policy. It attains a balance between accuracy and latency by dynamically selecting the optimal pruning level and split point. Experimental results across various tasks demonstrate that Janus enhances throughput by up to 5.15 times and reduces latency violation ratios by up to 98.7% when compared with baseline approaches under various network environments.

Figures

Figures reproduced from arXiv: 2502.10047 by the authors.

**Figure 2.** (a) shows the communication latency required for uploading a compressed image through 4G, 5G, and WiFi. Not surprisingly, the 4G and 5G connections have slower performance, requiring 166.84 ms and 80.46 ms, respectively, while WiFi connections exhibit significantly reduced upload times, with 32.17 ms [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** System overview of Janus. approach may incur communication costs when data needs to be transmitted to the cloud, it compensates by reducing deviceside computation latency, leading to a decrease in total latency. Unlike CNNs, vanilla ViTs exhibit distinct structural characteristics. In the transformer architecture, the output data size remains relatively consistent. For instance, when considering ViT-B, a typical m… view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: The fine-to-coarse candidate splitting points generating [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Layer latency of ViTs across different numbers of [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Devices in our real-world deployment. ViT model is then prepared and executed. Its intermediate output is compressed using standard LZW compression before transmission to the Jcloud. In Jcloud, it initially receives information about the model type, split point, and de…

**Figure 7.** Figure 7: The overall performance under different network conditions and tasks. Janus shows a slight improvement in average [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

**Figure 8.** Figure 8: Illustration of how Janus works for the image recog [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

**Figure 9.** Figure 9: Latency and the corresponding decision results under varying bandwidths. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 46 canonical work pages

[1]

A survey of vision-based traffic monitoring of road intersections,

S. R. E. Datondji, Y . Dupuis, P. Subirats, and P. Vasseur, “A survey of vision-based traffic monitoring of road intersections,” IEEE Trans. Intell. Transp. Syst. , vol. 17, no. 10, pp. 2681–2698, 2016

work page 2016
[2]

A Guide to Video Analytics: Applications and Opportuni- ties,

Tryolabs, “A Guide to Video Analytics: Applications and Opportuni- ties,” https://tryolabs.com/guides/video-analytics-guide, (accessed Dec. 2, 2023)

work page 2023
[3]

Video Analytics in Practical AI Applications [2024 Guide],

N. Klingler, “Video Analytics in Practical AI Applications [2024 Guide],” https://viso.ai/computer-vision/video-analytics-ultimate- overview/, (accessed Dec. 2, 2023)

work page 2024
[4]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, and S. Gelly, “An image is worth 16x16 words: Transformers for image recognition at scale,” in Proc. ICLR, 2021

work page 2021
[5]

Exploring plain vision transformer backbones for object detection,

Y . Li, H. Mao, R. Girshick, and K. He, “Exploring plain vision transformer backbones for object detection,” in Proc. ECCV. Springer, 2022, pp. 280–296

work page 2022
[6]

Semantic segmentation using Vision Transformers: A survey,

H. Thisanke, C. Deshan, K. Chamith, S. Seneviratne, R. Vidanaarachchi, and D. Herath, “Semantic segmentation using Vision Transformers: A survey,” Eng. Appl. Artif. Intell. , vol. 126, p. 106669, 2023

work page 2023
[7]

Videomae v2: Scaling video masked autoencoders with dual masking,

L. Wang, B. Huang, Z. Zhao, Z. Tong, Y . He, Y . Wang, Y . Wang, and Y . Qiao, “Videomae v2: Scaling video masked autoencoders with dual masking,” in Proc. IEEE/CVF CVPR , 2023, pp. 14 549–14 560

work page 2023
[8]

Transformers in Vision: A Survey,

S. Khan, M. Naseer, M. Hayat, S. W. Zamir, F. S. Khan, and M. Shah, “Transformers in Vision: A Survey,” ACM Comput. Surv. , vol. 54, no. 10s, pp. 1–41, 2022

work page 2022

Show all 47 references

[9]

A survey on vision transformer,

K. Han, Y . Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y . Tang, A. Xiao, C. Xu, and Y . Xu, “A survey on vision transformer,”IEEE Trans. Pattern Anal. Mach. Intell. , vol. 45, no. 1, pp. 87–110, 2022

2022
[10]

MMClassification,

L. Yang, X. Li, Z. Lou, M. Yang, F. Wang, C. Qian, K. Chen, and D. Lin, “MMClassification,” https://github.com/open-mmlab/mmclassification, 2020

2020
[11]

Learning Efficient Vision Transformers via Fine-Grained Manifold Distillation,

Z. Hao, J. Guo, D. Jia, K. Han, Y . Tang, C. Zhang, H. Hu, and Y . Wang, “Learning Efficient Vision Transformers via Fine-Grained Manifold Distillation,” Proc. NeurIPS, vol. 35, pp. 9164–9175, 2022

2022
[12]

SA ViT: Structure-Aware Vision Transformer Pruning via Collaborative Optimization,

C. Zheng, Z. Li, K. Zhang, Z. Yang, W. Tan, J. Xiao, Y . Ren, and S. Pu, “SA ViT: Structure-Aware Vision Transformer Pruning via Collaborative Optimization,” Proc. NeurIPS, vol. 35, pp. 9010–9023, 2022

2022
[13]

PTQ4ViT: Post-training Quantization for Vision Transformers with Twin Uniform Quantization,

Z. Yuan, C. Xue, Y . Chen, Q. Wu, and G. Sun, “PTQ4ViT: Post-training Quantization for Vision Transformers with Twin Uniform Quantization,” in Proc. ECCV. Springer, 2022, pp. 191–207

2022
[14]

ShiftAddNAS: Hardware- inspired search for more accurate and efficient neural networks,

H. You, B. Li, S. Huihong, Y . Fu, and Y . Lin, “ShiftAddNAS: Hardware- inspired search for more accurate and efficient neural networks,” inProc. PMLR ICML, 2022, pp. 25 566–25 580

2022
[15]

MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer,

S. Mehta and M. Rastegari, “MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer,” in Proc. ICLR, 2022

2022
[16]

Decomposable Intelligence on Cloud-Edge IoT Framework for Live Video Analytics,

Y . Zhang, J.-H. Liu, C.-Y . Wang, and H.-Y . Wei, “Decomposable Intelligence on Cloud-Edge IoT Framework for Live Video Analytics,” IEEE Internet of Things J. , vol. 7, no. 9, pp. 8860–8873, 2020

2020
[17]

SPINN: Synergistic progressive inference of neural networks over device and cloud,

S. Laskaridis, S. I. Venieris, M. Almeida, I. Leontiadis, and N. D. Lane, “SPINN: Synergistic progressive inference of neural networks over device and cloud,” in Proc. ACM MobiCom , 2020, pp. 1–15

2020
[18]

Neurosurgeon: Collaborative Intelligence Between the Cloud and Mobile Edge,

Y . Kang, J. Hauswald, C. Gao, A. Rovinski, T. Mudge, J. Mars, and L. Tang, “Neurosurgeon: Collaborative Intelligence Between the Cloud and Mobile Edge,” in Proc. ACM ASPLOS , 2017, pp. 615–629

2017
[19]

CNNPC: End- Edge-Cloud Collaborative CNN Inference With Joint Model Partition and Compression,

S. Yang, Z. Zhang, C. Zhao, X. Song, S. Guo, and H. Li, “CNNPC: End- Edge-Cloud Collaborative CNN Inference With Joint Model Partition and Compression,” IEEE Trans. Parallel Distrib. Syst. , vol. 33, no. 12, pp. 4039–4056, 2022

2022
[20]

Lzw data compression,

M. R. Nelson, “Lzw data compression,” Dr. Dobb’s J., vol. 14, no. 10, pp. 29–36, 1989

1989
[21]

Ofcom Publish Benchmarks of UK 5G, 4G and 3G Mobile Networks,

M. Jackson, “Ofcom Publish Benchmarks of UK 5G, 4G and 3G Mobile Networks,” https://www.ispreview.co.uk/index.php/2023/07/ofcom- publish-benchmarks-of-uk-5g-4g-and-3g-mobile-networks.html, (accessed Nov. 10, 2023)

2023
[22]

An in-depth study of uplink performance of 5G mmWave networks,

M. Ghoshal, Z. J. Kong, Q. Xu, Z. Lu, S. Aggarwal, I. Khan, Y . Li, Y . C. Hu, and D. Koutsonikolas, “An in-depth study of uplink performance of 5G mmWave networks,” in Proc. ACM SIGCOMM Workshop 5G Measurements Model. Use Cases (5G-MeMU) , 2022, pp. 29–35

2022
[23]

Imagenet: A large-scale hierarchical image database,

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in IEEE CVPR , 2009, pp. 248–255

2009
[24]

Distilled Split Deep Neural Networks for Edge-Assisted Real-Time Systems,

Y . Matsubara, S. Baidya, D. Callegaro, M. Levorato, and S. Singh, “Distilled Split Deep Neural Networks for Edge-Assisted Real-Time Systems,” in Proc. ACM HotEdgeVideo, 2019, pp. 21–26

2019
[25]

A Reconfigurable Neural Ar- chitecture for Edge–Cloud Collaborative Real-Time Object Detection,

J. C. Lee, Y . Kim, S. Moon, and J. H. Ko, “A Reconfigurable Neural Ar- chitecture for Edge–Cloud Collaborative Real-Time Object Detection,” IEEE Internet of Things J. , vol. 9, no. 23, pp. 23 390–23 404, 2022

2022
[26]

Adaptive Token Sampling for Efficient Vision Transformers,

M. Fayyaz, S. A. Koohpayegani, F. R. Jafari, S. Sengupta, H. R. V . Joze, E. Sommerlade, H. Pirsiavash, and J. Gall, “Adaptive Token Sampling for Efficient Vision Transformers,” in Proc. ECCV. Springer, 2022, pp. 396–414

2022
[27]

Token Merging: Your ViT But Faster,

D. Bolya, C.-Y . Fu, X. Dai, P. Zhang, C. Feichtenhofer, and J. Hoffman, “Token Merging: Your ViT But Faster,” in Proc. ICLR, 2023

2023
[28]

Improving fairness, efficiency, and stability in http-based adaptive video streaming with festive,

J. Jiang, V . Sekar, and H. Zhang, “Improving fairness, efficiency, and stability in http-based adaptive video streaming with festive,” in Proc. ACM CoNEXT, 2012, pp. 97–108

2012
[29]

OpenCV - Open Computer Vision Library,

“OpenCV - Open Computer Vision Library,” https://opencv.org/, (ac- cessed Apr. 10, 2024)

2024
[30]

Pytorch Image Models (timm) | timmdocs,

“Pytorch Image Models (timm) | timmdocs,” https://timm.fast.ai/, (ac- cessed Apr. 10, 2024)

2024
[31]

NVIDIA Jetson Orin,

“NVIDIA Jetson Orin,” https://www.nvidia.com/en-us/autonomous- machines/embedded-systems/jetson-orin/, (accessed Apr. 10, 2024)

2024
[32]

Elastic Compute Service,

“Elastic Compute Service,” https://www.alibabacloud.com/help/en/ecs/, (accessed Apr. 10, 2024)

2024
[33]

The kinetics human action video dataset,

W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijaya- narasimhan, F. Viola, T. Green, T. Back, P. Natsev et al., “The kinetics human action video dataset,” arXiv preprint arXiv:1705.06950 , 2017

2017 arXiv
[34]

Masked Autoencoders As Spatiotemporal Learners,

C. Feichtenhofer, H. Fan, Y . Li, and K. He, “Masked Autoencoders As Spatiotemporal Learners,” Proc. NeurIPS , vol. 35, pp. 35 946–35 958, 2022

2022
[35]

IA-RED ˆ2: Interpretability-Aware Redundancy Reduction for Vision Transformers,

B. Pan, R. Panda, Y . Jiang, Z. Wang, R. Feris, and A. Oliva, “IA-RED ˆ2: Interpretability-Aware Redundancy Reduction for Vision Transformers,” Proc. NeurIPS, vol. 34, pp. 24 898–24 911, 2021

2021
[36]

Patch slimming for efficient vision transformers,

Y . Tang, K. Han, Y . Wang, C. Xu, J. Guo, C. Xu, and D. Tao, “Patch slimming for efficient vision transformers,” in Proc. IEEE/CVF CVPR , 2022, pp. 12 165–12 174

2022
[37]

Evo-vit: Slow-fast token evolution for dynamic vision transformer,

Y . Xu, Z. Zhang, M. Zhang, K. Sheng, K. Li, W. Dong, L. Zhang, C. Xu, and X. Sun, “Evo-vit: Slow-fast token evolution for dynamic vision transformer,” in Proc. AAAI, vol. 36, 2022, pp. 2964–2972

2022
[38]

Width & depth pruning for vision transformers,

F. Yu, K. Huang, M. Wang, Y . Cheng, W. Chu, and L. Cui, “Width & depth pruning for vision transformers,” in Proc. AAAI , vol. 36, 2022, pp. 3143–3151

2022
[39]

PSAQ-ViT V2: Toward Accurate and General Data-Free Quantization for Vision Transformers,

Z. Li, M. Chen, J. Xiao, and Q. Gu, “PSAQ-ViT V2: Toward Accurate and General Data-Free Quantization for Vision Transformers,” IEEE Trans. Neural Netw. Learn. Syst. , pp. 1–12, 2023

2023
[40]

Lite vision transformer with enhanced self-attention,

C. Yang, Y . Wang, J. Zhang, H. Zhang, Z. Wei, Z. Lin, and A. Yuille, “Lite vision transformer with enhanced self-attention,” in Proc. IEEE/CVF CVPR , 2022, pp. 11 998–12 008

2022
[41]

In-situ ai: Towards autonomous and incremental deep learning for iot systems,

M. Song, K. Zhong, J. Zhang, Y . Hu, D. Liu, W. Zhang, J. Wang, and T. Li, “In-situ ai: Towards autonomous and incremental deep learning for iot systems,” in Proc. IEEE HPCA , 2018, pp. 92–103

2018
[42]

Computation Offload- ing for Machine Learning Web Apps in the Edge Server Environment,

H.-J. Jeong, I. Jeong, H.-J. Lee, and S.-M. Moon, “Computation Offload- ing for Machine Learning Web Apps in the Edge Server Environment,” in Proc. IEEE ICDCS , 2018, pp. 1492–1499

2018
[43]

Dynamic adaptive DNN surgery for inference acceleration on the edge,

C. Hu, W. Bao, D. Wang, and F. Liu, “Dynamic adaptive DNN surgery for inference acceleration on the edge,” in Proc. IEEE INFOCOM, 2019, pp. 1423–1431

2019
[44]

Boomerang: On-Demand Cooperative Deep Neural Network Inference for Edge Intelligence on the Industrial Internet of Things,

L. Zeng, E. Li, Z. Zhou, and X. Chen, “Boomerang: On-Demand Cooperative Deep Neural Network Inference for Edge Intelligence on the Industrial Internet of Things,” IEEE Netw., vol. 33, no. 5, pp. 96–103, 2019

2019
[45]

Edge AI: On-Demand Accelerat- ing Deep Neural Network Inference via Edge Computing,

E. Li, L. Zeng, Z. Zhou, and X. Chen, “Edge AI: On-Demand Accelerat- ing Deep Neural Network Inference via Edge Computing,” IEEE Trans. Wirel. Commun., vol. 19, no. 1, pp. 447–457, 2020

2020
[46]

Fine-grained elastic partitioning for distributed dnn towards mobile web ar services in the 5g era,

P. Ren, X. Qiao, Y . Huang, L. Liu, C. Pu, and S. Dustdar, “Fine-grained elastic partitioning for distributed dnn towards mobile web ar services in the 5g era,” IEEE Trans. Serv. Comput. , vol. 15, no. 6, pp. 3260–3274, 2022

2022
[47]

Towards cloud- edge collaborative online video analytics with fine-grained serverless pipelines,

M. Zhang, F. Wang, Y . Zhu, J. Liu, and Z. Wang, “Towards cloud- edge collaborative online video analytics with fine-grained serverless pipelines,” in Proc. ACM MMSys , 2021, pp. 80–93

2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

A survey of vision-based traffic monitoring of road intersections,

S. R. E. Datondji, Y . Dupuis, P. Subirats, and P. Vasseur, “A survey of vision-based traffic monitoring of road intersections,” IEEE Trans. Intell. Transp. Syst. , vol. 17, no. 10, pp. 2681–2698, 2016

work page 2016

[2] [2]

A Guide to Video Analytics: Applications and Opportuni- ties,

Tryolabs, “A Guide to Video Analytics: Applications and Opportuni- ties,” https://tryolabs.com/guides/video-analytics-guide, (accessed Dec. 2, 2023)

work page 2023

[3] [3]

Video Analytics in Practical AI Applications [2024 Guide],

N. Klingler, “Video Analytics in Practical AI Applications [2024 Guide],” https://viso.ai/computer-vision/video-analytics-ultimate- overview/, (accessed Dec. 2, 2023)

work page 2024

[4] [4]

An image is worth 16x16 words: Transformers for image recognition at scale,

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, and S. Gelly, “An image is worth 16x16 words: Transformers for image recognition at scale,” in Proc. ICLR, 2021

work page 2021

[5] [5]

Exploring plain vision transformer backbones for object detection,

Y . Li, H. Mao, R. Girshick, and K. He, “Exploring plain vision transformer backbones for object detection,” in Proc. ECCV. Springer, 2022, pp. 280–296

work page 2022

[6] [6]

Semantic segmentation using Vision Transformers: A survey,

H. Thisanke, C. Deshan, K. Chamith, S. Seneviratne, R. Vidanaarachchi, and D. Herath, “Semantic segmentation using Vision Transformers: A survey,” Eng. Appl. Artif. Intell. , vol. 126, p. 106669, 2023

work page 2023

[7] [7]

Videomae v2: Scaling video masked autoencoders with dual masking,

L. Wang, B. Huang, Z. Zhao, Z. Tong, Y . He, Y . Wang, Y . Wang, and Y . Qiao, “Videomae v2: Scaling video masked autoencoders with dual masking,” in Proc. IEEE/CVF CVPR , 2023, pp. 14 549–14 560

work page 2023

[8] [8]

Transformers in Vision: A Survey,

S. Khan, M. Naseer, M. Hayat, S. W. Zamir, F. S. Khan, and M. Shah, “Transformers in Vision: A Survey,” ACM Comput. Surv. , vol. 54, no. 10s, pp. 1–41, 2022

work page 2022