Pith. sign in

REVIEW 4 major objections 5 minor 58 references

DetVPCC: RoI-based Point Cloud Sequence Compression for 3D Object Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read DetVPCC improves VPCC for 3D object detection by compressing object regions at higher quality and background at lower quality.

desk verdict Solid RoI-aware VPCC extension with plausible gains, but the headline AmAP values are tuned on the test split and need a held-out check before I'd trust the exact numbers. read the letter →

arxiv 2502.04804 v1 pith:KWGFZBZA submitted 2025-02-07 cs.CV

Mingxuan Yan , Ruijie Zhang , Xuedou Xiao , Wei Wang This is my paper · ORCID
classification cs.CV
keywords 3DobjectdetectionpointcloudcompressionVPCCregion-of-interestencodingLiDARsequencebitrate-accuracytrade-offheatmap-basedRoInon-uniformqualityassignment
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

Video-based point cloud compression (VPCC) spends its bits uniformly across a scene, so heavy compression blurs exactly the small objects a 3D detector needs. This paper proposes DetVPCC, which first finds likely object regions with a lightweight heatmap detector and then compresses the corresponding depth-image macroblocks at a higher quality than background regions. On a large autonomous-driving dataset, DetVPCC reports an averaged mean-average-precision advantage over vanilla VPCC of up to 3.61 points with one common LiDAR detector and 2.80 with another, across a range of bitrates. The claim matters because it suggests a standard-compatible point cloud codec can be made machine-aware without redesigning the back-end detector.

What carries the argument

The mechanism is non-uniform quantization: H.264's macroblock-level QP control lets the encoder preserve fine geometric detail in object regions while heavily quantizing the rest of the depth image, and since geometry images dominate the VPCC bitstream, this is where the bitrate savings come from. The RoI map is produced by a GMM-based heatmap detector: each object's points are modeled as a mixture of five 3D Gaussians, projected to a 2D heatmap, binarized at threshold 0.4, unioned across classes, and masked by a ground-removal foreground map, then mapped from point-level to macroblock-level via the VPCC point-to-pixel map. The network is deliberately lighter than a full object detector because it does not need class labels or box regression, only object-likelihood heatmaps.

What would settle it

Take a detector that uses surrounding context, such as road layout, free space, or object-to-object relations, and run it on DetVPCC-encoded sequences at a high background QP versus uniform VPCC at matched bitrate; if the context-aware detector's mAP advantage over vanilla VPCC falls well below the 3.61 and 2.80 averages reported here, or turns negative, the RoI-equals-objects assumption is the reason.

Watch

Extended reading notes

Core claim

The central claim is that the poor bitrate-accuracy trade-off of VPCC under 3D object detection is largely a problem of where bits are spent, and that a binary region-of-interest split is enough to fix it. DetVPCC transcodes VPCC's lossless geometry (depth) images with H.264, assigning a low QP to macroblocks inside predicted RoIs and a high QP to background macroblocks. The RoI mask comes from a lightweight detector that outputs a class-union heatmap over a pillar grid, trained on ground-truth boxes via Gaussian mixture models rather than full bounding-box regression. The paper reports that this scheme consistently improves mAP, NDS, and most true-positive metrics over uniform VPCC for two back-end detectors across three RoI QPs and six background QPs, with the largest gains on small objects such as pedestrians and barriers.

Load-bearing premise

The method assumes that the regions a 3D detector needs in high detail are exactly the regions where objects sit, so background compression does not hurt detection accuracy.

Editorial extensions

If this is right

  • At matched bitrate, a stream encoded with DetVPCC should let a downstream 3D detector find small objects that uniform VPCC smears away, especially pedestrians, barriers, and bicycles.
  • A given detection-accuracy target can be reached at lower bitrate, which matters for streaming LiDAR to cloud servers or archiving large autonomous-driving datasets.
  • Because only the transcoding step changes, the method can attach to an existing VPCC pipeline and existing pre-trained detectors fine-tuned on a mix of lossy and lossless frames.
  • The RoI detector's overhead is small relative to the back-end detectors, with under 50 GFLOPs and under five million parameters, so the gain does not depend on an expensive preprocessing stage.

Reading between the lines

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

  • If a detector relies on context outside object boxes, such as road edges, free space, or the arrangement of static scene elements, the binary RoI split could discard information those detectors need; testing on context-sensitive detectors would likely bracket the size of the reported gains.
  • The paper fixes a single RoI QP per sequence; allowing the RoI/background QP split to vary frame-by-frame with object density or scene complexity is a natural extension that could shift the reported bitrate-accuracy curves further.
  • The same RoI principle could transfer to other point-cloud tasks, such as segmentation, tracking, or retrieval, by redefining the importance heatmap; the paper mentions broader applications only briefly.
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 / 5 minor

Summary. The paper proposes DetVPCC, a point cloud sequence compression framework that extends MPEG VPCC with region-of-interest (RoI) encoding for machine vision. A lightweight GMM-based heatmap detector predicts RoIs in each LiDAR frame; these RoIs are mapped to macroblock-level quality variations in VPCC's depth images, with lower QP inside RoIs and higher QP in the background. The method is evaluated on nuScenes with two back-end 3D detectors (CenterPoint and BEVFusion-Lidar), and the headline claim is that DetVPCC consistently improves the bitrate-accuracy trade-off of vanilla VPCC, with averaged mAP advantages of up to 3.61 and 2.80 (Table 2). The paper also includes ablations on the number of GMM components K, the binarization threshold gamma, and ground removal, plus a computational speed test for the point-in-box labeling algorithm.

Significance. If the reported gains are reliable, DetVPCC is a useful practical contribution: it is, to my knowledge, the first work to explicitly address the poor bitrate-accuracy trade-off of VPCC when the decompressed point cloud is consumed by a 3D object detector, and it does so within the standard VPCC framework using only macroblock-level QP control. The method is conceptually simple, the code is promised in the supplementary material, the comparison against a Naive RoI baseline is reasonable, and the use of two different back-end detectors is a strength. The significance is conditional on the evaluation being trustworthy: the headline numbers come from a single split with hyperparameters that appear to be tuned on the test set, so the magnitude of the claimed advantage is not yet established with confidence.

major comments (4)
  1. [Sec. 4.1.4, Sec. 4.3.2, Sec. 4.3.3] The values K=5 and gamma=0.4 are fixed in Sec. 4.1.4, but the ablations in Tables 4 and 5 are evaluated on the same split that later serves as the test set (the original nuScenes validation set, Sec. 4.1.1). Because Table 2's headline AmAP values (3.61 for CenterPoint and 2.80 for BEVFusion-Lidar) are produced with these fixed values, the hyperparameters appear to be selected after observing test-set performance, which can inflate the measured advantage. Please tune K and gamma on a disjoint validation split drawn from the training scenes, then freeze them before reporting the main comparison; the same concern applies to the statement in Sec. 4.2 that 'we empirically keep qr no larger than 25', which is justified by inspecting the test-set curves.
  2. [Sec. 4.1.1 and Table 2] The evaluation is restricted to a 40-frame clip per scene (frames 90 to 130) and all reported numbers come from a single run without error bars, confidence intervals, or significance tests. The claim that 'DetVPCC consistently improves VPCC' (Sec. 4.2) is therefore not backed by any uncertainty quantification. Please provide variability estimates, for example by bootstrapping over scenes or running multiple detector fine-tunes, and report whether the AmAP differences in Table 2 are outside the noise level.
  3. [Sec. 5 (Limitation 1) and Suppl. 8.1] The authors acknowledge in Sec. 5 that the back-end detectors are fine-tuned on uniformly compressed point clouds, not on the RoI-compressed data that the deployed system actually produces. This is a load-bearing mismatch: the reported accuracy is obtained from a model that has not seen the DetVPCC output distribution, so the measured AmAP may understate or misrepresent the bitrate-accuracy trade-off of the full pipeline. Please fine-tune the back-end detectors on RoI-compressed data (or on a mixture that includes it) and report whether the advantage over vanilla VPCC changes.
  4. [Sec. 5 (Limitation 3) and Sec. 4.2] The paper assumes that RoIs are exactly where objects are located, while acknowledging that non-object regions may contain information critical for object identification. Since the binary RoI/background quality split discards such context, the claim that DetVPCC 'consistently improves' across back-end detectors is only tested on CenterPoint and BEVFusion-Lidar, both of which may be relatively insensitive to context. Please test a context-sensitive detector or a soft/context-aware RoI variant to determine whether the gains persist when the assumption is relaxed.
minor comments (5)
  1. [Table 3] There is a typo in the class name 'pedestrain' (should be 'pedestrian'), and the em-dash entries in the class-wise table are not explained; please state what they mean or use a consistent notation for missing values.
  2. [Eq. (3)] The notation in Eq. (3) is unclear: the dimensions and operations involving Mp, Mb, R[:,t], and Rt T are not defined precisely. Please rewrite the equation with explicit matrix dimensions and explain what each variable represents.
  3. [Eq. (7)] The averaged advantage A is defined as an integral over the overlapping domain of the two metric-bitrate curves. Since the curves of different methods may have different domains, please state explicitly how xmin and xmax are chosen and whether the results are sensitive to the interpolation procedure and to the choice of the six qb values.
  4. [References] The reference for ScanNet is listed as 'Dai et al.' without a full author list or venue details; please complete the citation.
  5. [Sec. 4.3.2] The explanation for the decline at K=7 (that a simpler representation stabilizes learning) is plausible but not directly evidenced; a sentence or figure showing training convergence or heatmap quality for different K would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DetVPCC's reported gains are empirical comparisons, not derived from fitted inputs or self-citation chains.

full rationale

The paper's load-bearing claim is that RoI encoding improves the bitrate-accuracy trade-off of VPCC. The derivation chain is: (1) train a lightweight GMM-based RoI detector on ground-truth bounding boxes, (2) binarize its heatmap into a RoI mask, (3) apply non-uniform QP assignment in H.264 transcoding of VPCC depth images, and (4) measure the resulting mAP/NDS advantage against vanilla VPCC with back-end detectors that were fine-tuned on a mixture of lossless and uniformly compressed point clouds, not on RoI-compressed data. The advantage metric A in Eq. (7) is a measured integral of metric-bitrate curves, not a quantity that is forced to be positive by construction. No equation in the paper equates the claimed output (improved detection accuracy) to the method's input (the RoI detector's training labels or the chosen QPs). The choices of K=5 and gamma=0.4 are empirical, and the paper reports ablations over them; tuning hyperparameters on the same split is a statistical validity concern, not structural circularity, and should be assessed under correctness risk rather than circularity. The only self-citation, reference [48], appears in related work and is not load-bearing for the method's validity. The central comparison is an external, code-reproducible empirical evaluation against a standard VPCC baseline, so the paper is self-contained against external benchmarks.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities; its load-bearing elements are a GMM heatmap supervision scheme and two tuned hyperparameters, plus several domain assumptions about VPCC behavior and object-centric importance.

free parameters (2)
  • K, number of Gaussian mixture components per object = 5
    Chosen by ablation on the evaluation split (Table 4); K=5 gives the best AmAP.
  • gamma, heatmap binarization threshold = 0.4
    Chosen by ablation on the evaluation split (Table 5); gamma=0.4 gives best AmAP and NDS.
assumptions (4)
  • ad hoc to paper Object point clouds are drawn from a 3D Gaussian mixture with uniform component weights (Sec. 3.3.1, Eq. 4).
    This is a modeling simplification introduced for the RoI heatmap; it is not forced by the data and the paper does not validate the GMM fit against alternative distributions.
  • domain assumption The NVENC emphasis map implements macroblock-level QP control as intended on VPCC depth images (Sec. 3.2.2).
    The method relies on a proprietary driver feature to realize non-uniform quality; the paper gives no verification of the actual QP map on decoded images.
  • domain assumption Ground-truth bounding boxes correctly mark the regions whose fidelity matters for back-end detection (Sec. 3.3, Sec. 5 limitation 3).
    The authors themselves flag that non-object regions may carry critical information, so this assumption is load-bearing and acknowledged as a limitation.
  • domain assumption Back-end detectors fine-tuned on uniformly compressed point clouds still benefit from RoI-compressed test data (Sec. 5 limitation 1).
    The evaluation compares a detector that saw vanilla-compressed data during fine-tuning against RoI-compressed data at test time; the transfer is assumed but not directly optimized.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DetVPCC: RoI-based Point Cloud Sequence Compression for 3D Object Detection." pith.science (2026). https://pith.science/paper/KWGFZBZA

@misc{pith2026250204804,
  author       = {Pith},
  title        = {Pith review of: DetVPCC: RoI-based Point Cloud Sequence Compression for 3D Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KWGFZBZA}},
  note         = {Machine review of arXiv:2502.04804}
}
read the original abstract

While MPEG-standardized video-based point cloud compression (VPCC) achieves high compression efficiency for human perception, it struggles with a poor trade-off between bitrate savings and detection accuracy when supporting 3D object detectors. This limitation stems from VPCC's inability to prioritize regions of different importance within point clouds. To address this issue, we propose DetVPCC, a novel method integrating region-of-interest (RoI) encoding with VPCC for efficient point cloud sequence compression while preserving the 3D object detection accuracy. Specifically, we augment VPCC to support RoI-based compression by assigning spatially non-uniform quality levels. Then, we introduce a lightweight RoI detector to identify crucial regions that potentially contain objects. Experiments on the nuScenes dataset demonstrate that our approach significantly improves the detection accuracy. The code and demo video are available in supplementary materials.

Figures

Figures reproduced from arXiv: 2502.04804 by the authors.

Figure 1
Figure 1. The first row demonstrates the accuracy improvement of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of DetVPCC. application of compression may result in the loss of critical details in RoIs, potentially compromising the functionality of 3D object detectors. To better understand this dilemma, [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Illustration of RoI Encoding.2 The right image compares point clouds with RoI and non-RoI encoding. The red points rep￾resent the points that appear in the non-RoI encoded frame but not in the RoI encoded frame, and blue represents the contrary. Black points represent common parts. where E is a scaling matrix, and ⊗ represents element-wise multiplication. W = CXCT is the unscaled coefficient matrix, which is quantiz… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Design of the GMM-based RoI detector. accuracy function; qr is the pre-defined QP of RoI; the non￾RoI (background) QP qb is sampled from a pre-defined QP value set Qb to comprehensively evaluate R under different background point cloud qualities. Typically, RoIs are de…
Figure 5
Figure 5. Figure 5: Illustration of the evaluation metric.4 dataset described in Sec. 4.1.1. Ground removal and RoI encoding follow the same procedure as DetVPCC. To fairly compare the GMM-based RoI detector and naive RoI detector, the network capacity of the two de￾tectors is comparable,…
Figure 6
Figure 6. Figure 6: Qualitative results. The first and second rows are the detection results of CenterPoint on point clouds encoded with DetVPCC [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Metric-bitrate curve in Sec. 4.2 (CenterPoint) [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Metric-bitrate curve in Sec. 4.2 (BEVFusion-Lidar) 9. Metric-bitrate Curves in Main Results [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 55 canonical work pages

  1. [1]

    https : / / www

    nuscenes.org. https : / / www . nuscenes . org / object-detection. [Accessed 29-10-2024]. 5

    work page 2024

  2. [2]

    Inter- frame compression for dynamic point cloud geometry cod- ing

    Anique Akhtar, Zhu Li, and Geert Van der Auwera. Inter- frame compression for dynamic point cloud geometry cod- ing. IEEE Transactions on Image Processing, 2024. 2

    work page 2024

  3. [3]

    Zamir, Helen Jiang, Ioan- nis Brilakis, Martin Fischer, and Silvio Savarese

    Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioan- nis Brilakis, Martin Fischer, and Silvio Savarese. 3d seman- tic parsing of large-scale indoor spaces. In Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition, 2016. 8

    work page 2016

  4. [4]

    The quickhull algorithm for convex hulls.ACM Transactions on Mathematical Software (TOMS), 22(4):469–483, 1996

    C Bradford Barber, David P Dobkin, and Hannu Huhdanpaa. The quickhull algorithm for convex hulls.ACM Transactions on Mathematical Software (TOMS), 22(4):469–483, 1996. 5

    work page 1996

  5. [5]

    3D people surveillance on range data se- quences of a rotating Lidar

    Csaba Benedek. 3D people surveillance on range data se- quences of a rotating Lidar. Pattern Recognition Letters, 50: 149–158, 2014. 1

    work page 2014

  6. [6]

    Lidar-assisted 3d human detection for video surveillance

    Miquel Romero Blanch, Zenjie Li, Sergio Escalera, and Ka- mal Nasrollahi. Lidar-assisted 3d human detection for video surveillance. In Proceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision, pages 123–131,

    work page

  7. [7]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuScenes: A Multi- modal Dataset for Autonomous Driving. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11618–11628, Seattle, W A, USA, 2020. IEEE. 1, 2, 5, 6

    work page 2020

  8. [8]

    MMDetection3D: Open- MMLab next-generation platform for general 3D object detection

    MMDetection3D Contributors. MMDetection3D: Open- MMLab next-generation platform for general 3D object detection. https://github.com/open- mmlab/ mmdetection3d, 2020. 5, 1

    work page 2020

Show all 58 references
  1. [9]

    Motion- compensated compression of dynamic voxelized point clouds

    Ricardo L De Queiroz and Philip A Chou. Motion- compensated compression of dynamic voxelized point clouds. IEEE Transactions on Image Processing , 26(8): 3886–3895, 2017. 2

    2017

  2. [10]

    Server-driven video streaming for deep learning inference

    Kuntai Du, Ahsan Pervaiz, Xin Yuan, Aakanksha Chowdh- ery, Qizheng Zhang, Henry Hoffmann, and Junchen Jiang. Server-driven video streaming for deep learning inference. In Proceedings of the Annual conference of the ACM Special Interest Group on Data Communication on the applic...

    2020

  3. [11]

    Accmpeg: Optimiz- ing video encoding for video analytics

    Kuntai Du, Qizheng Zhang, Anton Arapin, Haodong Wang, Zhengxu Xia, and Junchen Jiang. Accmpeg: Optimiz- ing video encoding for video analytics. arXiv preprint arXiv:2204.12534, 2022. 2

    2022 arXiv

  4. [12]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    Dai et al. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, IEEE, 2017. 8

    2017

  5. [13]

    Point-cloud based 3D object detection and classification methods for self-driving applications: A survey and taxonomy

    Duarte Fernandes, Ant ´onio Silva, Rafael N ´evoa, Cl ´audia Sim˜oes, Dibet Gonzalez, Miguel Guevara, Paulo Novais, Jo˜ao Monteiro, and Pedro Melo-Pinto. Point-cloud based 3D object detection and classification methods for self-driving applications: A survey and taxonomy. Info...

    2021

  6. [14]

    Geometry coding for dynamic vox- elized point clouds using octrees and multiple contexts.IEEE Transactions on Image Processing, 29:313–322, 2019

    Diogo C Garcia, Tiago A Fonseca, Renan U Ferreira, and Ricardo L De Queiroz. Geometry coding for dynamic vox- elized point clouds using octrees and multiple contexts.IEEE Transactions on Image Processing, 29:313–322, 2019. 2

    2019

  7. [15]

    Graph-based network for dynamic point cloud prediction

    Pedro Gomes. Graph-based network for dynamic point cloud prediction. In Proceedings of the 12th ACM Multimedia Sys- tems Conference, pages 393–397, 2021. 2

    2021

  8. [16]

    Graziosi, O

    D. Graziosi, O. Nakagami, S. Kuma, A. Zaghetto, T. Suzuki, and A. Tabatabai. An overview of ongoing point cloud com- pression standardization activities: video-based (V-PCC) and geometry-based (G-PCC). APSIPA Transactions on Sig- nal and Information Processing, 9(1), 2020. 1, 2, 3

    2020

  9. [17]

    Video-based dynamic mesh cod- ing

    Danillo Bracco Graziosi. Video-based dynamic mesh cod- ing. In 2021 IEEE International Conference on Image Pro- cessing (ICIP), pages 3133–3137. IEEE, 2021. 1

    2021

  10. [18]

    Discover 3d surveillance, 2021

    Hexagon. Discover 3d surveillance, 2021. 1

    2021

  11. [19]

    Himmelsbach, Felix v

    M. Himmelsbach, Felix v. Hundelshausen, and H.-J. Wuen- sche. Fast segmentation of 3D point clouds for ground ve- hicles. In 2010 IEEE Intelligent Vehicles Symposium, pages 560–565, 2010. ISSN: 1931-0587. 1

    2010

  12. [20]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 1

    2014 arXiv

  13. [21]

    Pointpillars: Fast encoders for object detection from point clouds

    Alex H Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697–12705, 2019. 5

    2019

  14. [22]

    Cornernet: Detecting objects as paired keypoints

    Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European confer- ence on computer vision (ECCV) , pages 734–750, 2018. 4, 5

    2018

  15. [23]

    Patch- work++: Fast and robust ground segmentation solving par- tial under-segmentation using 3D point cloud

    Seungjae Lee, Hyungtae Lim, and Hyun Myung. Patch- work++: Fast and robust ground segmentation solving par- tial under-segmentation using 3D point cloud. In Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. , pages 13276– 13283, 2022. 5

    2022

  16. [24]

    Patch- work++: Fast and Robust Ground Segmentation Solving Par- tial Under-Segmentation Using 3D Point Cloud

    Seungjae Lee, Hyungtae Lim, and Hyun Myung. Patch- work++: Fast and Robust Ground Segmentation Solving Par- tial Under-Segmentation Using 3D Point Cloud. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 13276–13283, 2022. ISSN: 2153-0866. 1

    2022

  17. [25]

    Task-driven semantic cod- ing via reinforcement learning

    Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic cod- ing via reinforcement learning. IEEE Transactions on Image Processing, 30:6307–6320, 2021. 2

    2021

  18. [26]

    3d object detection for aerial plat- forms via edge computing: An experimental evaluation

    Alexander Lianides, Isaac Chan, Mohamed Ismail, Ian Harshbarger, Marco Levorato, Davide Callegaro, and Sharon LG Contreras. 3d object detection for aerial plat- forms via edge computing: An experimental evaluation. In 2022 18th International Conference on Distributed Com- puti...

    2022

  19. [27]

    Pchm-net: A new point cloud compression framework for both human vision and machine vision

    Lei Liu, Zhihao Hu, and Jing Zhang. Pchm-net: A new point cloud compression framework for both human vision and machine vision. In 2023 IEEE International Conference on Multimedia and Expo (ICME) , pages 1997–2002. IEEE,

    2023

  20. [28]

    Adamask: Enabling machine-centric video streaming with adaptive frame mask- ing for dnn inference offloading

    Shengzhong Liu, Tianshi Wang, Jinyang Li, Dachun Sun, Mani Srivastava, and Tarek Abdelzaher. Adamask: Enabling machine-centric video streaming with adaptive frame mask- ing for dnn inference offloading. In Proceedings of the 30th ACM international conference on multimedia , pa...

    2022

  21. [29]

    Bevfusion: Multi- task multi-sensor fusion with unified bird’s-eye view repre- sentation

    Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi- task multi-sensor fusion with unified bird’s-eye view repre- sentation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774–2781. I...

    2023

  22. [30]

    3D Object Detection for Autonomous Driving: A Comprehensive Survey

    Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hong- sheng Li. 3D Object Detection for Autonomous Driving: A Comprehensive Survey. International Journal of Computer Vision, 131(8):1909–1963, 2023. 1

    1909

  23. [31]

    Towards edge-assisted real-time 3d segmenta- tion of large scale lidar point clouds

    Fraser McLean, Leyang Xue, Chris Xiaoxuan Lu, and Ma- hesh Marina. Towards edge-assisted real-time 3d segmenta- tion of large scale lidar point clouds. In Proceedings of the 6th International Workshop on Embedded and Mobile Deep Learning, pages 1–6, 2022. 1

    2022

  24. [32]

    Performance assessment of point cloud compression

    Rufael Mekuria, Sebastien Laserre, and Christian Tulvan. Performance assessment of point cloud compression. In 2017 IEEE Visual Communications and Image Processing (VCIP), pages 1–4, 2017. 1

    2017

  25. [33]

    Dao: Dy- namic adaptive offloading for video analytics

    Taslim Murad, Anh Nguyen, and Zhisheng Yan. Dao: Dy- namic adaptive offloading for video analytics. In Proceed- ings of the 30th ACM International Conference on Multime- dia, pages 3017–3025, 2022. 2

    2022

  26. [34]

    NVENC Video Encoder API - Emphasis MAP,

    Nvidia. NVENC Video Encoder API - Emphasis MAP,

  27. [35]

    Frustum-pointpillars: A multi-stage ap- proach for 3d object detection using rgb camera and lidar

    Anshul Paigwar, David Sierra-Gonzalez, ¨Ozg¨ur Erkent, and Christian Laugier. Frustum-pointpillars: A multi-stage ap- proach for 3d object detection using rgb camera and lidar. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2926–2933, 2021. 1

    2021

  28. [36]

    Frustum pointnets for 3d object detection from rgb- d data

    Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb- d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918–927, 2018. 1

    2018

  29. [37]

    Deep video codec control for vision models

    Christoph Reich, Biplob Debnath, Deep Patel, Tim Prange- meier, Daniel Cremers, and Srimat Chakradhar. Deep video codec control for vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5732–5741, 2024. 2

    2024

  30. [38]

    Garden Richardson

    Iain E. Garden Richardson. H. 264/mpeg-4 part 10 white peper : Transform & quantization. 2003. 3

    2003

  31. [39]

    Rabbit: Live transcoding of v-pcc point cloud streams

    Michael Rudolph, Stefan Schneegass, and Amr Rizk. Rabbit: Live transcoding of v-pcc point cloud streams. In Proceed- ings of the 14th Conference on ACM Multimedia Systems , pages 97–107, 2023. 3

    2023

  32. [40]

    Continual learning for real-world autonomous systems: Algorithms, challenges and frameworks

    Khadija Shaheen, Muhammad Abdullah Hanif, Osman Hasan, and Muhammad Shafique. Continual learning for real-world autonomous systems: Algorithms, challenges and frameworks. Journal of Intelligent & Robotic Systems , 105 (1):9, 2022. 1

    2022

  33. [41]

    A rate control algorithm for video-based point cloud compression

    Fangyu Shen and Wei Gao. A rate control algorithm for video-based point cloud compression. In 2021 International Conference on Visual Communications and Image Process- ing (VCIP), pages 1–5. IEEE, 2021. 3

    2021

  34. [42]

    An advanced lidar point cloud sequence cod- ing scheme for autonomous driving

    Xuebin Sun, Sukai Wang, Miaohui Wang, Shing Shin Cheng, and Ming Liu. An advanced lidar point cloud sequence cod- ing scheme for autonomous driving. In Proceedings of the 28th ACM International Conference on Multimedia , pages 2793–2801, 2020. 2

    2020

  35. [43]

    Geometric distortion metrics for point cloud compression

    Dong Tian, Hideaki Ochimizu, Chen Feng, Robert Cohen, and Anthony Vetro. Geometric distortion metrics for point cloud compression. In 2017 IEEE International Conference on Image Processing (ICIP), pages 3460–3464, 2017. ISSN: 2381-8549. 1

    2017

  36. [44]

    Vqba: Visual-quality-driven bit allocation for low- latency point cloud streaming

    Shuoqian Wang, Mufeng Zhu, Na Li, Mengbai Xiao, and Yao Liu. Vqba: Visual-quality-driven bit allocation for low- latency point cloud streaming. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9143– 9151, 2023. 3

    2023

  37. [45]

    Enabling edge-cloud video analytics for robotics applications

    Yiding Wang, Weiyan Wang, Duowen Liu, Xin Jin, Junchen Jiang, and Kai Chen. Enabling edge-cloud video analytics for robotics applications. IEEE Transactions on Cloud Com- puting, 11(2):1500–1513, 2022. 2

    2022

  38. [46]

    Overview of the h

    Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding stan- dard. IEEE Transactions on circuits and systems for video technology, 13(7):560–576, 2003. 1, 3

    2003

  39. [47]

    3D-LIDAR Based Object De- tection and Tracking on the Edge of IoT for Railway Level Crossing

    Cristian Wisultschew, Gabriel Mujica, Jose Manuel Lanza- Gutierrez, and Jorge Portilla. 3D-LIDAR Based Object De- tection and Tracking on the Edge of IoT for Railway Level Crossing. IEEE Access, 9:35718–35729, 2021. Conference Name: IEEE Access. 1

    2021

  40. [48]

    Dnn-driven compressive offloading for edge-assisted semantic video segmentation

    Xuedou Xiao, Juecheng Zhang, Wei Wang, Jianhua He, and Qian Zhang. Dnn-driven compressive offloading for edge-assisted semantic video segmentation. In IEEE IN- FOCOM 2022-IEEE Conference on Computer Communica- tions, pages 1888–1897. IEEE, 2022. 2

    2022

  41. [49]

    Source compression with bounded dnn perception loss for iot edge computer vision

    Xiufeng Xie and Kyu-Han Kim. Source compression with bounded dnn perception loss for iot edge computer vision. In The 25th Annual International Conference on Mobile Com- puting and Networking, pages 1–16, 2019. 2

    2019

  42. [50]

    Online continual learning for 3D detection of road participants in autonomous driving

    Rui Yang. Online continual learning for 3D detection of road participants in autonomous driving . PhD thesis, Universit ´e Bourgogne Franche-Comt´e, 2023. 1

    2023

  43. [51]

    Center- based 3d object detection and tracking

    Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center- based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784–11793, 2021. 1, 4, 5, 7

    2021

  44. [52]

    RT3D: Real-Time 3-D Vehi- cle Detection in LiDAR Point Cloud for Autonomous Driv- ing

    Yiming Zeng, Yu Hu, Shice Liu, Jing Ye, Yinhe Han, Xi- aowei Li, and Ninghui Sun. RT3D: Real-Time 3-D Vehi- cle Detection in LiDAR Point Cloud for Autonomous Driv- ing. IEEE Robotics and Automation Letters , 3(4):3434– 3440, 2018. Conference Name: IEEE Robotics and Automa- tio...

    2018

  45. [53]

    Casva: Configuration-adaptive streaming for live video analytics

    Miao Zhang, Fangxin Wang, and Jiangchuan Liu. Casva: Configuration-adaptive streaming for live video analytics. In IEEE INFOCOM 2022-IEEE Conference on Computer Communications, pages 2168–2177. IEEE, 2022. 2

    2022

  46. [54]

    Ada3d: Exploiting the spatial redun- dancy with adaptive inference for efficient 3d object detec- tion

    Tianchen Zhao, Xuefei Ning, Ke Hong, Zhongyuan Qiu, Pu Lu, Yali Zhao, Linfeng Zhang, Lipu Zhou, Guohao Dai, Huazhong Yang, et al. Ada3d: Exploiting the spatial redun- dancy with adaptive inference for efficient 3d object detec- tion. In Proceedings of the IEEE/CVF Internationa...

    2023

  47. [55]

    Ob- jects as points

    Xingyi Zhou, Dequan Wang, and Philipp Kr ¨ahenb¨uhl. Ob- jects as points. arXiv preprint arXiv:1904.07850, 2019. 4, 7 DetVPCC: RoI-based Point Cloud Sequence Compression for 3D Object Detection Supplementary Material

    1904 arXiv

  48. [56]

    Header Design Inspired by CenterPoint [51], we separate the heatmap pre- dictions of objects with different sizes into 6 tasks

    Details of RoI detectors 7.1. Header Design Inspired by CenterPoint [51], we separate the heatmap pre- dictions of objects with different sizes into 6 tasks. Back- bone features first pass through a shared convolution block, which includes a 3 × 3 convolution layer, a batch no...

  49. [57]

    Training We pre-train the back-end detectors with lossless point clouds following the settings in MMDetection3D [8]

    Details of Back-end 3D Object Detectors 8.1. Training We pre-train the back-end detectors with lossless point clouds following the settings in MMDetection3D [8]. Specifically, we train each detector for 20 epochs using the Adam optimizer. The batch size is set to 9, the learni...

  50. [58]

    7 and Fig

    Metric-bitrate Curves in Main Results Fig. 7 and Fig. 8 give the mAP-bitrate and NDS-bitrate curves of DetVPCC, VPCC and Naive RoI in Sec. 4.2

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.