Pith. sign in

REVIEW 4 major objections 6 minor 20 references

QuantV2X: A Fully Quantized Multi-Agent System for Cooperative Perception

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read QuantV2X claims that a fully quantized cooperative perception system—low-bit models plus code-index messages—matches full-precision accuracy and, once real transmission latency is counted, beats full-precision systems by 3.2× in end-to-end

desk verdict Model-level PTQ work is solid; the headline system-level gains rest on a simulation the paper never specifies. read the letter →

arxiv 2509.03704 v2 pith:VLTBL4SN submitted 2025-09-03 cs.CV

Seth Z. Zhao , Huizhi Zhang , Zhaowei Li , Juntong Peng , Anthony Chui , Zewei Zhou , Zonglin Meng , Hao Xiang
show 6 more authors
Zhiyu Huang Fujia Wang Ran Tian Chenfeng Xu Bolei Zhou Jiaqi Ma
This is my paper
classification cs.CV
keywords V2Xcooperativeperceptionpost-trainingquantizationcodebookrepresentationBEVfeaturefusionsystem-levellatencymulti-agentsystemsautonomousdrivingintermediate
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

QuantV2X tries to establish that the real obstacle to V2X cooperative perception in the field is not detection accuracy but system-level efficiency: full-precision models and full-precision feature messages are too slow, too large, and too memory-hungry for real-time multi-agent deployment. The paper claims that quantizing the whole pipeline—weights, activations, and the messages agents exchange—can preserve detection accuracy (up to 99.8% of full-precision AP under INT4/INT8) while cutting end-to-end latency by 3.2× and improving mAP30 by 9.5 points over the full-precision system once transmission delays are counted. The reason to care: most prior cooperative-perception work is evaluated in an idealized synchronized setting that ignores deployment cost; if the claim holds, quantization is what makes large, capable perception models actually fit inside vehicle GPUs and communication budgets.

What carries the argument

The load-bearing object is the shared codebook: a learnable dictionary of feature vectors that all agents use to quantize their BEV features, such that inter-agent messages carry only the nearest-code indices (reconstruction via a weighted combination of selected codes). Around it, the method stacks (i) post-training quantization of weights and activations with block-wise reconstruction and learnable rounding, calibrated on 0.5% of training data, and (ii) an alignment module in the fusion stage applying two losses—KL divergence between full-precision and quantized fused features (heterogeneity alignment) and L2 distance between their bounding-box outputs (spatial alignment)—that counter feat

What would settle it

On V2X-Real, re-run the system-level comparison with an explicit protocol: any feature message whose computed arrival time (feature size / 27 Mbps + uniform 0–200 ms jitter) exceeds the fusion deadline is dropped, and mAP30 is recomputed. If the full-precision system's gap to the upper bound (currently −10.7) does not widen far beyond QuantV2X's (−1.2) under this protocol, the deployment-oriented advantage evaporates.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a shared learnable codebook plus end-to-end post-training quantization turns intermediate-fusion cooperative perception into a deployable system without an accuracy tax. Each agent no longer transmits FP32 bird's-eye-view features; it sends only codebook indices (roughly 0.03 MB instead of 8.6 MB), and the receiver reconstructs features locally. Because stage-aware calibration also aligns quantized fused features and detection outputs with their full-precision counterparts via heterogeneity and spatial alignment losses, low-bit inference recovers essentially full-precision accuracy. The paper's central experimental claim is that when comparison

Load-bearing premise

The system-level comparison assumes that the simulation's mapping from communication delay to degraded detection accuracy—the paper never says whether late feature messages are dropped, delayed, or reordered—faithfully reproduces how real V2X networks lose information, and the headline +9.5 mAP30 gain hangs entirely on that unstated mapping.

Editorial extensions

If this is right

  • Fully quantized intermediate fusion is a viable deployment route: INT4-weight/INT8-activation models keep up to 99.8% of full-precision AP30, and the fully INT8 system matches full precision on V2X-Real and OPV2V.
  • Counting communication delay changes the ranking: QuantV2X's 3.2× system-level latency reduction turns into a +9.5 mAP30 advantage over full-precision baselines on V2X-Real (52.6 vs 43.1 mAP30).
  • Codebook-index messaging reduces per-message transmission from 8.6 MB (FP32 BEV features) to 0.03 MB, shrinking communication latency by roughly 5.3×.
  • Quantization rescales memory-constrained model selection, letting larger backbones run inside fixed vehicle-GPU budgets where only down-sized full-precision models otherwise fit.
  • The recipe generalizes across computation-based, CNN-based, and attention-based fusion methods—with the vision-transformer fusion method as the clear exception, losing about 27 points of AP30 at INT8.

Reading between the lines

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

  • The paper's headline +9.5 mAP30 is determined by a latency-to-accuracy protocol that is never specified (how delayed feature messages are handled once communication time pushes past the fusion deadline); re-running the system-level comparison under an explicit drop-stale-messages rule would tell how broadly the result holds.
  • The shared-codebook index trick should transfer to any bandwidth-limited multi-agent perception setting—robot swarms, distributed camera networks, drone teams—where agents exchange learned features rather than raw observations.
  • Since the vision-transformer fusion method collapses under low bits, architecture choice may matter more than calibration effort: designing fusion architectures from the start with low-bit-friendly operators could be a more robust path than post-hoc alignment.
  • For a fixed GPU memory budget, the accuracy-versus-efficiency tradeoff may invert: a quantized large model could beat a full-precision small model outright; the paper's scaling results support this, and a direct same-budget head-to-head would make it a crisp claim.
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.

Referee Report

4 major / 6 minor

Summary. The paper proposes QuantV2X, a fully quantized multi-agent cooperative perception system for V2X scenarios. The system has three stages: full-precision pretraining of an intermediate fusion model, codebook learning for quantized communication, and post-training quantization of the full model. A calibration procedure with block-wise reconstruction and two alignment losses (heterogeneity KL divergence and spatial L2) is introduced to reduce quantization-induced degradation. The authors present model-level results showing small accuracy drops under INT4/8 quantization across multiple fusion methods, and system-level results on V2X-Real claiming a 3.2x latency reduction and +9.5 mAP30 over full-precision baselines, plus scaling experiments under memory budgets.

Significance. If the system-level claims are reproducible, this is a useful contribution: it is among the first attempts to quantize the full stack of intermediate-fusion cooperative perception, including model inference and message representation, and it provides real TensorRT/ROS latency and power measurements, ablations of calibration size/steps, and intent to release code. The model-level PTQ results, especially on Pyramid Fusion, are credible and well-ablated. However, the headline system-level result currently rests on an unspecified latency-to-accuracy simulation, and the term 'fully quantized' is not fully supported as written. I do not see circular reasoning; the concern is verifiability and internal consistency of the system-level evaluation.

major comments (4)
  1. [Section 3.2.2 (Table 3)] The paper defines T_comm = f_s/v + U(0,200) and Tsys = Tlocal + Tcomm + Tfus, but it never specifies the mechanism by which Tsys degrades detection accuracy. No algorithm or equation describes whether delayed messages are dropped, replaced with stale features from past frames, reweighted, or otherwise injected at fusion time. The +9.5 mAP30 claim in the abstract cannot be reproduced or falsified without this protocol. For example, with f_s=8.6 MB and v=27 Mbps, the No Compression row has T_comm roughly 2.5 s; the resulting 43.1 mAP30 cannot be derived from the stated formulas. Please give the exact mapping (frame index selection, rounding, handling of delays longer than history, and sampling of U) and state explicitly whether the abstract's +9.5 is 52.6-43.1 (No Compression) or 52.6-48.8 (compressed full precision).
  2. [Section 3.2.1 vs 3.2.2] There is an inconsistency between the measured latency breakdown and the simulation formula. Figure 5 reports a total full-precision system latency of 127.7 ms, with Tcomm approximately 44.9 ms for the large BEV feature, but the formula in Section 3.2.2 with v=27 Mbps gives Tcomm = 8.6 MB * 8 / 27 Mbps = about 2.55 s for the same feature size. If Figure 5 is a real deployment measurement, then the simulated system in Table 3 is a different, much slower channel. If the formula is the intended simulation, then the measured breakdown is irrelevant. Please reconcile the two latency models or report the actual simulated Tcomm for every row of Table 3.
  3. [Section 2.3.1 (Codebook) and title claim] The transmitted messages are integer codebook indices, but the paper does not state whether the codebook entries D themselves are quantized. As written, D appears to be a set of FP32 learnable vectors, and the reconstructed feature is computed as a weighted sum of these FP32 vectors. Since the title and abstract claim a 'fully quantized' system, this is a load-bearing detail. Either quantize the codebook (and state the bit width) or explicitly state that the codebook table remains FP32 and argue why this does not undermine 'fully quantized'. Also, the dimension of the message-size formula in this section needs correction: bytes should be H*W*log2(nL)*nR/8, not log2(H*W*log2(nL)*nR/8).
  4. [Section 3.1.1 (Table 1) and Conclusion/Limitation] The claim that the PTQ method 'generalizes well across various fusion methods' is overstated in light of the V2X-ViT row of Table 1: under 4/8 W/A, AP30 drops from 57.4 to 29.9 and AP50 from 49.5 to 8.8; even 8/8 shows a large drop. The limitation section acknowledges significant degradation for V2X-ViT, but the main-text claim in Section 3.1.1 should be qualified to computation-based and CNN-based methods (and attention methods with simple structures). This is not a fatal flaw because the system-level system uses Pyramid Fusion, but it affects the stated contribution of a generalizable PTQ recipe.
minor comments (6)
  1. [Abstract and Section 1] The abstract says '+9.5 improvement in mAP30' and the introduction says '9.5% mAP30'. Express this as +9.5 mAP30 points to avoid percentage ambiguity, and identify the baseline (No Compression vs compressed) in the abstract.
  2. [Table 3] The row labels 'Full-Precision (Lu et al., 2024) BEV Feature/8.6 MB (No Compression)' and 'BEV Feature/0.54 MB (x16 Compression)' are ambiguous about which compressor is used and whether these rows have the same model weights as the Upper-bound. Please specify the compressor details and the full-precision training configuration for each row.
  3. [Figure 5] The numeric values in the figure are not self-explanatory; the caption does not list whether the bars are Tlocal, Tcomm, and Tfus and which color corresponds to which component. Report the exact per-component values in the caption or a table.
  4. [Table 10 vs Table 3] Table 10 reports for nC=128, nR=1 a system accuracy of 52.6/41.9, while Table 3 reports QuantV2X as 52.6/42.2 with message size 0.03 MB. Reconcile the discrepancy or state that Table 3 uses a slightly different configuration.
  5. [Appendix Table 6] The 'Low-Prec.' baseline in Table 6 is not defined. State whether it is FP16, bf16, or another quantization scheme, and give its bit configuration.
  6. [Section 3.2.2] Specify the units and sampling of U(0,200): is it in milliseconds, and is it sampled independently per agent, per frame, or once per evaluation? This is needed to reproduce the simulation.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; central results are empirical measurements, with a system-level simulation gap that is a verifiability issue rather than a circular reduction.

full rationale

The paper's chain is: (1) train an FP32 intermediate-fusion model; (2) learn a shared codebook and quantized indices (Eqs. 3-4); (3) apply PTQ with reconstruction and alignment losses (Eqs. 8-10); (4) measure model-level AP on held-out test sets and system-level latency on TensorRT/ROS (Fig. 5, Table 3). None of these steps defines its claimed output in terms of the output itself. The codebook reconstruction objective and PTQ calibration losses are optimization targets on calibration data; the reported AP/mAP numbers come from separate test evaluation, so they are not equal to those targets by construction. The system-level +9.5 mAP30 claim in the abstract and Table 3 depends on a latency-simulation protocol (Sec. 3.2.2, Tsys = Tlocal + Tcomm + Tfus with Tcomm = fs/v + U(0,200)) whose mapping from latency to accuracy is never specified; this is a real reproducibility/verifiability gap, but the paper borrows the protocol from prior work rather than fitting it to force the result, so it is not a circular reduction. Self-citations to the authors' V2X-Real dataset and V2X-ReaLO platform are dataset/tool references and are not used as mathematical evidence for the quantization or fusion claims. The limitation paragraph admits that V2X-ViT degrades under aggressive quantization, which is a negative result consistent with empirical measurement rather than a circular defense. Overall: no load-bearing circular step; score 2 only for the presence of minor non-load-bearing self-citations.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claims rely on several tuned hyperparameters (calibration search, codebook size, loss weights) and on an external latency simulation whose details are not fully specified. No new physical entities are introduced.

free parameters (5)
  • calibration scale search hyperparameters alpha, beta, T = not reported numerically
    Section 2.4.2: the optimal scale is chosen by minimizing reconstruction error over a range [alpha*s0, beta*s0]; alpha, beta, and T are not given.
  • reconstruction loss weight lambda_rec = not reported
    Eq. (4) balances detection loss and codebook reconstruction loss; the value is not stated.
  • codebook size nC and number of codes nR = nC=128, nR=1 used for main results
    Appendix D.3 and Table 10: configuration is selected based on highest system-level accuracy, so it is tuned on the evaluation metric.
  • calibration dataset fraction and steps = 0.5% of training data, 5000 steps
    Appendix C.1 and Tables 4/5: chosen after ablation on validation metrics.
  • transmission rate v and async delay U(0,200) = v=27 Mbps, U uniform 0-200 ms
    Section 3.2.2: taken from prior work (Arena and Pau, 2019; Xu et al., 2022b) and used in the system-level simulation.
assumptions (3)
  • domain assumption The PTQ pipeline (scale search, AdaRound, block-wise reconstruction) is a valid and accurate method for quantizing 3D detection models.
    The paper relies on standard PTQ techniques and only compares against LiDAR-PTQ and PD-Quant; the generalization of these methods to multi-agent fusion is assumed.
  • domain assumption The system-level latency model T_comm = f_s/v + U(0,200) and the way latency degrades detection accuracy (via delayed or dropped features) faithfully captures real-world deployment.
    Section 3.2.2: the simulation follows prior protocols, but the mapping from latency to mAP is not described in this paper, so the result depends on that external protocol's assumptions.
  • domain assumption The calibration dataset (0.5% of training) is representative of deployment conditions for quantization parameter estimation.
    Common PTQ assumption; validated only by a calibration-size ablation in Table 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QuantV2X: A Fully Quantized Multi-Agent System for Cooperative Perception." pith.science (2026). https://pith.science/paper/VLTBL4SN

@misc{pith2026250903704,
  author       = {Pith},
  title        = {Pith review of: QuantV2X: A Fully Quantized Multi-Agent System for Cooperative Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VLTBL4SN}},
  note         = {Machine review of arXiv:2509.03704}
}
abstract

Cooperative perception through Vehicle-to-Everything (V2X) communication offers significant potential for enhancing vehicle perception by mitigating occlusions and expanding the field of view. However, past research has predominantly focused on improving accuracy metrics without addressing the crucial system-level considerations of efficiency, latency, and real-world deployability. Noticeably, most existing systems rely on full-precision models, which incur high computational and transmission costs, making them impractical for real-time operation in resource-constrained environments. In this paper, we introduce \textbf{QuantV2X}, the first fully quantized multi-agent system designed specifically for efficient and scalable deployment of multi-modal, multi-agent V2X cooperative perception. QuantV2X introduces a unified end-to-end quantization strategy across both neural network models and transmitted message representations that simultaneously reduces computational load and transmission bandwidth. Remarkably, despite operating under low-bit constraints, QuantV2X achieves accuracy comparable to full-precision systems. More importantly, when evaluated under deployment-oriented metrics, QuantV2X reduces system-level latency by 3.2$\times$ and achieves a +9.5 improvement in mAP30 over full-precision baselines. Furthermore, QuantV2X scales more effectively, enabling larger and more capable models to fit within strict memory budgets. These results highlight the viability of a fully quantized multi-agent intermediate fusion system for real-world deployment. The system will be publicly released to promote research in this field: https://github.com/ucla-mobility/QuantV2X.

Figures

Figures reproduced from arXiv: 2509.03704 by the authors.

Figure 1
Figure 1. Motivation. Left: Full-precision cooperative perception systems are ill-suited for real-world deployment. Right: QuantV2X presents an efficient and scalable solution for real-world cooperative driving systems. To bridge this gap, we present QuantV2X, a fully quantized multi-agent cooperative system designed to holistically address the system-level latency and performance drop in resource-constrained V2X settings (sh… view at source ↗
Figure 2
Figure 2. QuantV2X overview. On the model side, QuantV2X transforms full-precision neural [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Effectiveness of the proposed alignment module. Compared to naive quantization, (a) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Robustness under pose errors (Collaboration mode: [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: System-level latency breakdown (unit: ms). Note that the numbers are obtained through [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Scaling trend under dif￾ferent GPU memory constraints. QuantV2X enables deployment of larger models under tight budgets while maintaining high perception performance. We examine the scaling behavior of QuantV2X by varying the backbone capacity of both the full-precisio…
Figure 7
Figure 7. Figure 7: Qualitative results on DAIR-V2X dataset (Collaboration mode: [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results on DAIR-V2X dataset (Collaboration mode: [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Illustration of real-world testing platform. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 7 canonical work pages

  1. [4]

    Gptq: Accurate post-training quantization for generative pre-trained transformers

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323,

    arXiv

  2. [7]

    CooperRisk: A driving risk quantification pipeline with multi-agent cooperative perception and prediction

    Mingyue Lei, Zewei Zhou, Hongchen Li, Jia Hu, and Jiaqi Ma. CooperRisk: A driving risk quantification pipeline with multi-agent cooperative perception and prediction. arXiv preprint arXiv:2506.15868, 2025a. Mingyue Lei, Zewei Zhou, Hongchen Li, Jiaqi Ma, and Jia Hu. Risk map as middleware: Towards interpretable cooperative end-to-end autonomous driving fo...

    arXiv

  3. [9]

    When2com: Multi-agent perception via communication graph grouping

    Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020a. Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learna...

    work page 2020

  4. [12]

    Analysis of v2x communication parameters for the development of a fusion architecture for cooperative perception systems

    Andreas Rauch, Felix Klanner, and Klaus Dietmayer. Analysis of v2x communication parameters for the development of a fusion architecture for cooperative perception systems. In 2011 IEEE Intelligent V ehicles Symposium (IV), pages 685–690,

    work page 2011

  5. [14]

    V2vnet: Vehicle-to-vehicle communication for joint perception and prediction

    Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16 , pages 605–621. Springer,

    work page 2020

  6. [15]

    Easyquant: Post-training quantization via scale optimization

    Yuhang Wu, Yunhe Wang, Kai Han, Chunjing Huang, and Qi Tian. Easyquant: Post-training quantization via scale optimization. arXiv preprint arXiv:2006.16669,

    arXiv 2006

  7. [16]

    V2x-real: a largs-scale dataset for vehicle-to-everything cooperative perception

    Hao Xiang, Zhaoliang Zheng, Xin Xia, Runsheng Xu, Letian Gao, Zewei Zhou, Xu Han, Xinkai Ji, Mingxi Li, Zonglin Meng, et al. V2x-real: a largs-scale dataset for vehicle-to-everything cooperative perception. arXiv preprint arXiv:2403.16034,

    arXiv

  8. [17]

    Cobevt: Cooperative bird’s eye view semantic segmentation with sparse transformers

    Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird’s eye view semantic segmentation with sparse transformers. arXiv preprint arXiv:2207.02202, 2022a. Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to- everything cooperative perception with vision transformer....

    arXiv 2022

Show all 20 references
  1. [18]

    Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization

    Li Yuan, Zhaohui Zhang, Shaopeng Hao, and Jiashi Feng. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. arXiv preprint arXiv:2111.12293,

    arXiv

  2. [19]

    Coopre: Cooperative pretraining for v2x cooperative perception

    Seth Z Zhao, Hao Xiang, Chenfeng Xu, Xin Xia, Bolei Zhou, and Jiaqi Ma. Coopre: Cooperative pretraining for v2x cooperative perception. arXiv preprint arXiv:2408.11241,

    arXiv

  3. [20]

    LiDAR-PTQ: Post-training quantization for point cloud 3d object detection

    Sifan Zhou, Liang Li, Xinyu Zhang, Bo Zhang, Shipeng Bai, Miao Sun, Ziyu Zhao, Xiaobo Lu, and Xiangxiang Chu. LiDAR-PTQ: Post-training quantization for point cloud 3d object detection. In The Twelfth International Conference on Learning Representations, 2024a. Zewei Zhou, Hao ...

    arXiv

  4. [2011]

    Collaborative automated driving: A machine learning-based method to enhance the accuracy of shared information

    Zaydoun Yahya Rawashdeh and Zheng Wang. Collaborative automated driving: A machine learning-based method to enhance the accuracy of shared information. In 2018 21st International Conference on Intelligent Transportation Systems (ITSC), pages 3961–3966. IEEE,

    2018

  5. [2017]

    Markus Nagel, Raoul Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort

    Presentation. Markus Nagel, Raoul Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. arXiv preprint arXiv:2004.10568,

    2004 arXiv

  6. [2019]

    Actnn: Reducing training memory footprint via 2-bit activation compressed training.arXiv preprint arXiv:2104.14129,

    Jianfei Chen, Jang-Hyun Choi, Xinyi Zhou, Dionysios Brand, Joseph E Gonzalez, and Ion Stoica. Actnn: Reducing training memory footprint via 2-bit activation compressed training.arXiv preprint arXiv:2104.14129,

    arXiv

  7. [2020]

    Loss aware post-training quantization

    Yaniv Nahshan, Ronen Banner, Itay Hubara, Elad Hoffer, Daniel Soudry, Alexander M Bronstein, and Avi Mendelson. Loss aware post-training quantization. arXiv preprint arXiv:1911.07190,

    1911 arXiv

  8. [2021]

    F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds

    Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing , pages 88–100, 2019a. Qi Chen, Sih...

    2019

  9. [2022]

    Yue Hu, Juntong Peng, Sifei Liu, Junhao Ge, Si Liu, and Siheng Chen

    Curran Associates Inc. Yue Hu, Juntong Peng, Sifei Liu, Junhao Ge, Si Liu, and Siheng Chen. Communication-Efficient Collaborative Perception via Information Filling with Codebook . In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15481–15490...

    2024

  10. [2023]

    Learned step size quantization

    Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153,

    1902 arXiv

  11. [2024]

    Ravi Krishnamoorthi

    IEEE Computer Society. Ravi Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342,

    arXiv

  12. [2025]

    Pd-quant: Post-training quantization based on prediction difference metric

    Ling Liu, Shuang Li, Yuhang Gong, Xishan Yang, Wei Zhang, and Xiuqiang Hu. Pd-quant: Post-training quantization based on prediction difference metric. arXiv preprint arXiv:2212.07048,

    arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.