pith. machine review for the scientific record. sign in

arxiv: 2605.12220 · v1 · submitted 2026-05-12 · 💻 cs.CV · cs.AI· cs.LG· cs.RO

Recognition: 2 theorem links

· Lean Theorem

TriBand-BEV: Real-Time LiDAR-Only 3D Pedestrian Detection via Height-Aware BEV and High-Resolution Feature Fusion

Mohammad Khoshkdahan , Alexey Vinel
Authors on Pith no claims yet

Pith reviewed 2026-05-13 06:35 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LGcs.RO
keywords LiDARBEV3D pedestrian detectionreal-timeKITTIheight bandsautonomous drivingoriented bounding boxes
0
0 comments X

The pith

A three-band height-aware BEV turns full LiDAR scans into 2D detection followed by 3D box reconstruction for real-time pedestrian spotting.

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

The paper shows that the entire 3D LiDAR point cloud can be projected into a compact 2D bird's-eye-view tensor split across three fixed height bands. This encoding lets a single 2D network with area attention and bidirectional fusion predict oriented boxes, after which a simple reconstruction step lifts the results back to 3D space while an interquartile-range filter removes outliers. On the KITTI test set the method reports 58.7 / 52.6 / 47.2 percent BEV average precision for easy / moderate / hard pedestrians at 49 frames per second on one consumer GPU, beating the prior Complex-YOLO baseline by 12.6, 7.5 and 3.1 points respectively. The pipeline therefore supplies fast, single-pass detection of cars, pedestrians and cyclists without 3D convolutions or multi-view fusion.

Core claim

TriBand-BEV maps every LiDAR point into a lightweight 2D BEV tensor using three fixed vertical height bands, recasts 3D detection as ordinary 2D oriented-box regression, and recovers final 3D boxes from the 2D outputs via an IQR-based outlier filter. The network combines deep area attention, a hierarchical bidirectional neck fusing P1-P4 scales, distribution focal learning for side offsets, and a rotated IoU loss; training adds mild vertical re-binning and reflectance jitter.

What carries the argument

The TriBand-BEV encoding that divides the 3D point cloud into three fixed height bands to produce a 2D tensor supporting fast 2D detection and subsequent geometric reconstruction.

If this is right

  • One forward pass detects cars, pedestrians and cyclists together.
  • Pedestrian BEV AP reaches 58.7 / 52.6 / 47.2 on KITTI easy / moderate / hard while sustaining 49 FPS on a single consumer GPU.
  • Qualitative scenes remain stable when pedestrians are partially hidden by other objects.
  • The compact pipeline requires no explicit 3D convolutions or cross-view fusion.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Height-band encodings of this style could be tried on radar or camera-LiDAR fusion pipelines to lower compute.
  • Fixed three-band spacing may need scene-adaptive adjustment on datasets with extreme height variation.
  • Replacing the IQR filter with a small learned lifting head could further improve reconstruction on noisy scans.
  • Real-time speed makes online monitoring of band statistics feasible for domain adaptation.

Load-bearing premise

Projecting points into three fixed height bands plus simple 2D-to-3D lifting preserves enough geometry for accurate pedestrian detection even under occlusion and varying point density.

What would settle it

Run the same network on a LiDAR dataset rich in heavily occluded or unusually tall pedestrians and measure whether BEV AP falls more than 10 points below the KITTI numbers while a comparable full-3D detector does not.

Figures

Figures reproduced from arXiv: 2605.12220 by Alexey Vinel, Mohammad Khoshkdahan.

Figure 1
Figure 1. Figure 1: Detection pipeline overview. The pre-processing [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Network overview. The detection pipeline from [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Internal components of each module used in the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Random crops from the training set before [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Mean BEV and 3D AP across distance ranges. Each [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: 3D and BEV precision–recall curves for easy/moderate/hard on all three classes. 7.2 Qualitative Analysis [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative detection results on validation scenes. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
read the original abstract

Safe autonomous agents and mobile robots need fast real time 3D perception, especially for vulnerable road users (VRUs) such as pedestrians. We introduce a new bird's eye view (BEV) encoding, which maps the full 3D LiDAR point cloud into a light-weight 2D BEV tensor with three height bands. We explicitly reformulate 3D detection as a 2D detection problem and then reconstruct 3D boxes from the BEV outputs. A single network detects cars, pedestrians, and cyclists in one pass. The backbone uses area attention at deep stages, a hierarchical bidirectional neck over P1 to P4 fuses context and detail, and the head predicts oriented boxes with distribution focal learning for side offsets and a rotated IoU loss. Training applies a small vertical re bin and a mild reflectance jitter in channel space to resist memorization. We use an interquartile range (IQR) filter to remove noisy and outlier LiDAR points during 3D reconstruction. On KITTI dataset, TriBand-BEV attains 58.7/52.6/47.2 pedestrian BEV AP(%) for easy, moderate, and hard at 49 FPS on a single consumer GPU, surpassing Complex-YOLO, with gains of +12.6%, +7.5%, and +3.1%. Qualitative scenes show stable detection under occlusion. The pipeline is compact and ready for real time robotic deployment. Our source code is publicly available on GitHub.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces TriBand-BEV, a LiDAR-only pipeline for real-time 3D detection of pedestrians (and jointly cars/cyclists). It encodes the full point cloud into a lightweight 2D BEV tensor via three fixed height bands, reformulates detection as a 2D oriented-box task with area attention and a hierarchical bidirectional neck, then lifts predictions to 3D boxes via IQR filtering and a small vertical re-bin. On KITTI it reports pedestrian BEV AP of 58.7/52.6/47.2 (easy/moderate/hard) at 49 FPS on a single consumer GPU, outperforming Complex-YOLO by +12.6/+7.5/+3.1 points; code is released on GitHub.

Significance. If the performance numbers are reproducible and the fixed-band assumption holds under occlusion and sparsity, the method offers a compact, real-time alternative to full 3D convolutions or multi-view fusion for VRU detection. The public code release is a clear strength that enables direct verification and extension.

major comments (3)
  1. [Abstract] Abstract: the headline claim of +12.6/+7.5/+3.1 point gains over Complex-YOLO rests on KITTI pedestrian BEV AP numbers, yet the manuscript supplies no train/val split, no error bars, and no statement of whether results are single-run or averaged; this directly affects whether the central empirical result can be trusted.
  2. [Method] Method (height-band encoding and reconstruction): the three fixed height bands plus IQR-based 2D-to-3D lift are presented as sufficient to recover accurate 3D boxes even for hard/occluded pedestrians, but no ablation on band count, thresholds, or IQR multiplier is given; without these, it is impossible to judge whether the reported hard-case AP (47.2) is robust or an artifact of the chosen discretization.
  3. [Experiments] Experiments: the vertical re-bin and reflectance jitter are described only qualitatively ('small' and 'mild') with no quantitative sensitivity analysis; because these are the only explicit regularizers mentioned, their omission leaves open whether post-processing rather than the network architecture drives the claimed FPS and accuracy.
minor comments (2)
  1. [Abstract] The abstract and method text introduce the 'TriBand-BEV tensor' without an accompanying equation or figure that defines its exact channel layout and bin boundaries.
  2. [Qualitative results] A few sentences in the reconstruction paragraph use ambiguous phrasing ('stable detection under occlusion') that would benefit from a quantitative metric or additional qualitative examples.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below, clarifying our approach and committing to revisions that strengthen the empirical claims and analysis.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline claim of +12.6/+7.5/+3.1 point gains over Complex-YOLO rests on KITTI pedestrian BEV AP numbers, yet the manuscript supplies no train/val split, no error bars, and no statement of whether results are single-run or averaged; this directly affects whether the central empirical result can be trusted.

    Authors: We agree that explicit details improve trust in the results. The reported numbers use the standard KITTI split (3712 training / 3769 validation samples) as defined by the benchmark and followed in prior works such as Complex-YOLO. All results are from a single training run. We will add this information to the abstract and experiments section in the revision. Error bars are not provided due to the computational cost of multiple full trainings on our hardware; however, the public GitHub code release enables direct reproduction and optional multi-run averaging by others. We have added a clarifying statement on single-run reporting. revision: partial

  2. Referee: [Method] Method (height-band encoding and reconstruction): the three fixed height bands plus IQR-based 2D-to-3D lift are presented as sufficient to recover accurate 3D boxes even for hard/occluded pedestrians, but no ablation on band count, thresholds, or IQR multiplier is given; without these, it is impossible to judge whether the reported hard-case AP (47.2) is robust or an artifact of the chosen discretization.

    Authors: We acknowledge that ablations would help demonstrate robustness. The three fixed bands are chosen to span typical pedestrian height distributions (approximately 0–0.5 m, 0.5–1.5 m, >1.5 m) while preserving a compact BEV tensor for real-time inference; the IQR filter is a standard outlier removal step. In the revision we will add an ablation table varying band count (2/3/4), height thresholds, and IQR multiplier values, showing that the reported hard-case performance remains stable and is not an artifact of the specific discretization. revision: yes

  3. Referee: [Experiments] Experiments: the vertical re-bin and reflectance jitter are described only qualitatively ('small' and 'mild') with no quantitative sensitivity analysis; because these are the only explicit regularizers mentioned, their omission leaves open whether post-processing rather than the network architecture drives the claimed FPS and accuracy.

    Authors: We agree that a quantitative sensitivity study is warranted. The vertical re-bin and reflectance jitter serve as lightweight regularizers to reduce overfitting on sparse LiDAR data. We will insert a new table in the experiments section that reports pedestrian AP and FPS across a range of re-bin sizes and jitter amplitudes. This analysis will show that the core network (area attention, bidirectional neck, rotated IoU loss) accounts for the majority of the accuracy and speed gains, while the regularizers provide modest but consistent improvements without driving the overall results. revision: yes

Circularity Check

0 steps flagged

Empirical architecture with no derivation chain

full rationale

The paper introduces an empirical pipeline that encodes LiDAR points into a three-band BEV tensor, reformulates 3D detection as 2D detection, and reconstructs boxes via IQR filtering. No equations or theorems are presented whose outputs are shown to equal their inputs by construction, nor are any load-bearing claims justified solely by self-citation. Reported numbers are direct benchmark results on KITTI rather than predictions derived from fitted parameters within the same model. The design choices (fixed bands, area attention, rotated IoU loss) remain independent of the final AP figures.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The method rests on the domain assumption that three fixed height bands suffice to capture pedestrian geometry from LiDAR and that 2D detection plus simple reconstruction recovers accurate 3D boxes. No new physical entities are postulated. The height-band thresholds and network hyperparameters are free parameters chosen during design.

free parameters (1)
  • height band thresholds
    Three vertical intervals chosen to separate ground, pedestrian torso, and head regions; exact cutoffs are not stated in the abstract but are required for the encoding.
axioms (1)
  • domain assumption LiDAR points can be losslessly projected into a 2D BEV tensor with three height channels without discarding critical 3D structure for detection
    Invoked when reformulating 3D detection as 2D detection
invented entities (1)
  • TriBand-BEV tensor no independent evidence
    purpose: Compact 2D representation of 3D point cloud for fast detection
    New encoding scheme introduced by the paper

pith-pipeline@v0.9.0 · 5593 in / 1542 out tokens · 67174 ms · 2026-05-13T06:35:42.520923+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

54 extracted references · 54 canonical work pages · 2 internal anchors

  1. [1]

    Javed Ahmad and Alessio Del Bue. 2023. mmfusion: Multimodal fusion for 3d objects detection. arXiv preprint arXiv:2311.04058 (2023)

    work page arXiv 2023

  2. [2]

    Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. 2022. Transfusion: Robust lidar-camera fusion for 3d object de- tection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 1090–1099

    work page 2022

  3. [3]

    Xiaozhi Chen, Kaustav Kundu, Yukun Zhu, Andrew G Berneshawi, Huimin Ma, Sanja Fidler, and Raquel Urtasun. 2015. 3d object proposals for accurate object class detection. Advances in neural information processing systems 28 (2015)

    work page 2015

  4. [4]

    Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. 2017. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE con- ference on Computer Vision and Pattern Recognition . 1907–1915

    work page 2017

  5. [5]

    Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. 2021. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence , Vol. 35. 1201–1209

    work page 2021

  6. [6]

    Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. 2022. Embracing single stride 3d object detector with sparse transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8458–8468

    work page 2022

  7. [7]

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. 2012. Are we ready for Au- tonomous Driving? The KITTI Vision Benchmark Suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 3354–3361

    work page 2012

  8. [8]

    Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. 2020. Epnet: Enhancing point features with image semantics for 3d object detection. In European confer- ence on computer vision . 35–52

    work page 2020

  9. [9]

    Rahima Khanam and Muhammad Hussain. 2024. Yolov11: An overview of the key architectural enhancements. arXiv preprint arXiv:2410.17725 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    Mohammad Khoshkdahan, Arman Akbari, Arash Akbari, and Xuan Zhang

    work page

  11. [11]

    In International Conference on Intelligent Transportation Systems (ITSC)

    Beyond Overall Accuracy: Pose-and Occlusion-driven Fairness Analysis in Pedestrian Detection for Autonomous Driving. In International Conference on Intelligent Transportation Systems (ITSC)

    work page

  12. [12]

    Mohammad Khoshkdahan, Nicholas Kjär, and Fabian B Flohr. 2025. Fair-ped: Fairness evaluation in pedestrian detection using clip. In 2025 IEEE Intelligent Vehicles Symposium (IV). 1504–1509

    work page 2025

  13. [13]

    Bo Li, Tianlei Zhang, and Tian Xia. 2016. Vehicle detection from 3d lidar using fully convolutional network. arXiv preprint arXiv:1608.07916 (2016)

    work page arXiv 2016

  14. [14]

    Peizhao Li, Pu Wang, Karl Berntorp, and Hongfu Liu. 2022. Exploiting tempo- ral relations on radar perception for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 17071–17080

    work page 2022

  15. [15]

    Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. 2022. Hdmapnet: An online hd map construction and evaluation framework. In 2022 International Conference on Robotics and Automation (ICRA) . 4628–4634

    work page 2022

  16. [16]

    Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. 2018. Deep contin- uous fusion for multi-sensor 3d object detection. In Proceedings of the European conference on computer vision (ECCV) . 641–656

    work page 2018

  17. [17]

    Biao Liu and Yanxin Wu. 2025. BFT3D: A Robust BEV Feature Transformation Module for Multisensor 3-D Object Detection. IEEE Sensors Journal 25, 15 (2025), 30175–30185

    work page 2025

  18. [18]

    Wenjie Luo, Bin Yang, and Raquel Urtasun. 2018. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolu- tional net. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. 3569–3577

    work page 2018

  19. [19]

    Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. 2021. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. In Proceedings of the IEEE/CVF international conference on com- puter vision. 2723–2732

    work page 2021

  20. [20]

    Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. 2021. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision . 3164– 3173

    work page 2021

  21. [21]

    Gregory P Meyer, Jake Charland, Shreyash Pandey, Ankit Laddha, Shivam Gau- tam, Carlos Vallespi-Gonzalez, and Carl K Wellington. 2020. Laserflow: Efficient and probabilistic object detection and motion forecasting. IEEE Robotics and Automation Letters 6, 2, 526–533

    work page 2020

  22. [22]

    Gregory P Meyer, Ankit Laddha, Eric Kee, Carlos Vallespi-Gonzalez, and Carl K Wellington. 2019. Lasernet: An efficient probabilistic 3d object detector for au- tonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 12677–12686

    work page 2019

  23. [23]

    Yujian Mo, Yan Wu, Junqiao Zhao, Jijun Wang, Yinghao Hu, and Jun Yan. 2025. Enhancing LiDAR Point Features with Foundation Model Priors for 3D Object Detection. arXiv preprint arXiv:2507.13899 (2025)

    work page arXiv 2025

  24. [24]

    Pha Nguyen, Kha Gia Quach, Chi Nhan Duong, Ngan Le, Xuan-Bac Nguyen, and Khoa Luu. 2022. Multi-camera multiple 3d object tracking on the move for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 2569–2578

    work page 2022

  25. [25]

    Anshul Paigwar, Özgür Erkent, David Sierra-Gonzalez, and Christian Laugier

    work page

  26. [26]

    In 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS)

    GndNet: Fast ground plane estimation and point cloud segmentation for autonomous vehicles. In 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS) . IEEE, 2150–2156

    work page 2020

  27. [27]

    Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 2021. 3d ob- ject detection with pointformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 7463–7472

    work page 2021

  28. [28]

    Su Pang, Daniel Morris, and Hayder Radha. 2020. CLOCs: Camera-LiDAR object candidates fusion for 3D object detection. In 2020 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS) . 10386–10393

    work page 2020

  29. [29]

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

    work page 2018

  30. [30]

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. 2017. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition . 652–660

    work page 2017

  31. [31]

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. 2017. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30

    work page 2017

  32. [32]

    Yongxin Shao, Zhetao Sun, Aihong Tan, and Tianhong Yan. 2023. Efficient three- dimensional point cloud object detection based on improved Complex-YOLO. Frontiers in Neurorobotics 17 (2023), 1092564

    work page 2023

  33. [33]

    Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. 2020. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10529–10538

    work page 2020

  34. [34]

    Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xi- aogang Wang, and Hongsheng Li. 2023. PV-RCNN++: Point-voxel feature set abstraction with local vector representation for 3D object detection. Interna- tional Journal of Computer Vision 131, 2, 531–551

    work page 2023

  35. [35]

    Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 2019. Pointrcnn: 3d ob- ject proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 770–779

    work page 2019

  36. [36]

    Weijing Shi and Raj Rajkumar. 2020. Point-gnn: Graph neural network for 3d object detection in a point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 1711–1719

    work page 2020

  37. [37]

    Martin Simony, Stefan Milzy, Karl Amendey, and Horst-Michael Gross. 2018. Complex-yolo: An euler-region-proposal for real-time 3d object detection on point clouds. In Proceedings of the European conference on computer vision (ECCV) workshops

    work page 2018

  38. [38]

    Vishwanath A Sindagi, Yin Zhou, and Oncel Tuzel. 2019. Mvx-net: Multimodal voxelnet for 3d object detection. In 2019 International Conference on Robotics and Automation (ICRA). 7276–7282

    work page 2019

  39. [39]

    Yunjie Tian, Qixiang Ye, and David Doermann. 2025. Yolov12: Attention-centric real-time object detectors. arXiv preprint arXiv:2502.12524 (2025)

    work page internal anchor Pith review arXiv 2025

  40. [40]

    Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. 2020. Pointpaint- ing: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 4604–4612

    work page 2020

  41. [41]

    Zhixin Wang and Kui Jia. 2019. Frustum convnet: Sliding frustums to aggre- gate local point-wise features for amodal 3d object detection. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . 1742–1749

    work page 2019

  42. [42]

    Hai Wu, Jinhao Deng, Chenglu Wen, Xin Li, Cheng Wang, and Jonathan Li. 2022. CasA: A cascade attention network for 3-D object detection from LiDAR point clouds. IEEE Transactions on Geoscience and Remote Sensing 60 (2022), 1–11

    work page 2022

  43. [43]

    Hai Wu, Chenglu Wen, Wei Li, Xin Li, Ruigang Yang, and Cheng Wang. 2023. Transformation-equivariant 3d object detection for autonomous driving. In Pro- ceedings of the AAAI Conference on Artificial Intelligence , Vol. 37. 2795–2802

    work page 2023

  44. [44]

    Hai Wu, Chenglu Wen, Shaoshuai Shi, Xin Li, and Cheng Wang. 2023. Vir- tual sparse convolution for multimodal 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 21653–21662

    work page 2023

  45. [45]

    Yutian Wu, Yueyu Wang, Shuwei Zhang, and Harutoshi Ogai. 2020. Deep 3D object detection networks using LiDAR data: A review. IEEE Sensors Journal 21, 2 (2020), 1152–1171

    work page 2020

  46. [46]

    Yan Yan, Yuxing Mao, and Bo Li. 2018. Second: Sparsely embedded convolutional detection. Sensors 18, 10, 3337

    work page 2018

  47. [47]

    Bin Yang, Wenjie Luo, and Raquel Urtasun. 2018. Pixor: Real-time 3d object detection from point clouds. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition . 7652–7660

    work page 2018

  48. [48]

    Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 2020. 3dssd: Point-based 3d sin- gle stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 11040–11048

    work page 2020

  49. [49]

    Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. 2019. Std: Sparse- to-dense 3d object detector for point cloud. In Proceedings of the IEEE/CVF inter- national conference on computer vision . 1951–1960

    work page 2019

  50. [50]

    Wei Ye, Qiming Xia, Hai Wu, Zhen Dong, Ruofei Zhong, Cheng Wang, and Chenglu Wen. 2025. Fade3D: Fast and Deployable 3D Object Detection for Au- tonomous Driving. IEEE Transactions on Intelligent Transportation Systems 26, 9 (2025), 12934–12946

    work page 2025

  51. [51]

    Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 2020. 3d-cvf: Gen- erating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In European conference on computer vision . 720–736

    work page 2020

  52. [52]

    Zhuochen Yu, Bijie Qiu, and Andy WH Khong. 2025. ViKIENet: Towards Ef- ficient 3D Object Detection with Virtual Key Instance Enhanced Network. In Proceedings of the Computer Vision and Pattern Recognition Conference . 11844– 11853

    work page 2025

  53. [53]

    Jie Zhou, Xin Tan, Zhiwen Shao, and Lizhuang Ma. 2019. FVNet: 3D front-view proposal generation for real-time object detection from point clouds. In 12th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI). 1–8

    work page 2019

  54. [54]

    Yin Zhou and Oncel Tuzel. 2018. Voxelnet: End-to-end learning for point cloud based 3d object detection. InProceedings of the IEEE conference on computer vision and pattern recognition. 4490–4499

    work page 2018