Pith. sign in

REVIEW 2 major objections 63 references

Vanilla Vision Transformers match state-of-the-art performance on large-scale automotive lidar point cloud segmentation using a custom tokenizer, lightweight decoder, and tailored augmentations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-06-28 22:33 UTC pith:AY6ZRKUP

load-bearing objection VaViT shows a flat ViT can hit competitive lidar segmentation numbers with a tokenizer plus light head and augmentations, but the vanilla claim needs the tokenizer details to hold up. the 2 major comments →

arxiv 2605.31177 v1 pith:AY6ZRKUP submitted 2026-05-29 cs.CV

Vanilla ViT for Automotive Point Cloud Semantic Segmentation

Gilles Puy , Nermin Samet , Alexandre Boulch , Spyros Gidaris , Tuan-Hung VU , Renaud Marlet This is my paper
classification cs.CV
keywords Vision TransformerPoint cloud semantic segmentationAutomotive lidarnuScenesSemanticKITTIWaymo Open DatasetVanilla ViT
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper shows that plain non-hierarchical Vision Transformers can handle semantic segmentation of large automotive lidar scenes. Standard approaches rely on U-Net designs that interleave convolutions with local or windowed attention. A specialized tokenizer, simple decoder head, and targeted data augmentations close the usual performance gap. Tests on nuScenes, SemanticKITTI, and Waymo Open Dataset confirm the method reaches or surpasses existing results while preserving ViT simplicity.

Core claim

We show how to effectively leverage vanilla, non-hierarchical ViTs for segmentation of large-scale automotive lidar scenes. We bridge the performance gap thanks to a carefully designed tokenizer, a lightweight decoder segmentation head, and tailored data augmentations. Our approach, VaViT for Vanilla ViT, matches or exceeds the performance of state-of-the-art methods while maintaining the simplicity of ViT architecture.

What carries the argument

The VaViT tokenizer that converts point clouds into tokens for a standard ViT backbone, paired with a lightweight decoder segmentation head.

Load-bearing premise

A tokenizer, lightweight decoder, and data augmentations together suffice to overcome the missing hierarchical and convolutional biases in non-hierarchical ViTs.

What would settle it

Evaluating the released VaViT model on the nuScenes validation split and finding its mean IoU at least 3 points below the current leading method.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Standard ViT backbones become viable for point cloud segmentation without added hierarchy or convolutions.
  • The same tokenizer and decoder design can be tested on other large-scale lidar datasets.
  • Architectural simplicity reduces the engineering effort needed for multimodal fusion with image or text transformers.
  • Training recipes that rely only on data augmentations become sufficient for competitive lidar segmentation.

Where Pith is reading between the lines

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

  • Inductive biases from convolutions appear less essential once tokenization and augmentation are tuned for outdoor scenes.
  • The method could be adapted to other point cloud tasks such as instance segmentation or object detection on the same datasets.
  • Unified ViT pipelines may simplify sensor fusion across lidar, camera, and radar in production automotive stacks.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 0 minor

Summary. The manuscript presents VaViT, a vanilla non-hierarchical Vision Transformer architecture for semantic segmentation of large-scale automotive LiDAR point clouds. It claims that a carefully designed tokenizer, lightweight decoder segmentation head, and tailored data augmentations enable performance that matches or exceeds state-of-the-art methods on the nuScenes, SemanticKITTI, and Waymo Open Dataset benchmarks while preserving the simplicity of the plain ViT backbone. Code and models are released.

Significance. If the central empirical claim holds and the architecture remains free of injected hierarchical or convolutional biases, the result would be significant: it would demonstrate that plain transformers can close the gap on point-cloud segmentation without the U-Net-style interleaving of convolutions and local attentions that currently dominate the field, supporting broader unification of transformer backbones across modalities. Open-sourcing strengthens reproducibility.

major comments (2)
  1. [Method section (tokenizer description) and Experiments (ablation tables)] The central claim that tokenizer + lightweight decoder + augmentations alone suffice to compensate for the absence of hierarchical or convolutional inductive biases is load-bearing. Explicit ablations are required that (i) isolate the tokenizer design (voxelization, local grouping, or fixed-window partitioning) to confirm it does not introduce locality or multi-scale structure before the ViT backbone and (ii) apply equivalent preprocessing and augmentations to hierarchical baselines; without these controls the assertion that the method is purely 'vanilla ViT' remains under-determined.
  2. [Abstract and §4 (results tables)] Quantitative support for the 'matches or exceeds' claim is absent from the abstract and must be verified against the reported tables; if the gains are driven primarily by augmentations rather than the transformer itself, the architectural conclusion is weakened.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and describe the revisions that will be incorporated.

read point-by-point responses

  1. Referee: [Method section (tokenizer description) and Experiments (ablation tables)] The central claim that tokenizer + lightweight decoder + augmentations alone suffice to compensate for the absence of hierarchical or convolutional inductive biases is load-bearing. Explicit ablations are required that (i) isolate the tokenizer design (voxelization, local grouping, or fixed-window partitioning) to confirm it does not introduce locality or multi-scale structure before the ViT backbone and (ii) apply equivalent preprocessing and augmentations to hierarchical baselines; without these controls the assertion that the method is purely 'vanilla ViT' remains under-determined.

    Authors: We agree that stronger isolation of the tokenizer's contribution is valuable. Our tokenizer performs simple fixed-window partitioning into tokens with no local grouping, multi-scale processing, or learned locality prior to the ViT encoder, as specified in the method section. In the revised manuscript we will add dedicated ablation tables that vary only the tokenizer parameters (voxel size, window size) while holding the non-hierarchical ViT backbone, decoder, and augmentations fixed; these will confirm that no hierarchical bias is introduced before the transformer layers. For point (ii), we will add a paragraph comparing the preprocessing and augmentation pipelines used by the cited hierarchical baselines to our own, noting that our augmentations follow standard practices in the automotive LiDAR literature. Full re-training of every baseline under identical conditions is beyond the scope of a revision, but the added discussion will clarify the controls that are feasible. revision: yes

  2. Referee: [Abstract and §4 (results tables)] Quantitative support for the 'matches or exceeds' claim is absent from the abstract and must be verified against the reported tables; if the gains are driven primarily by augmentations rather than the transformer itself, the architectural conclusion is weakened.

    Authors: We will revise the abstract to include explicit quantitative statements drawn from the tables in §4 (e.g., mIoU on nuScenes, SemanticKITTI, and Waymo). The manuscript already contains component-wise ablations that separate the contributions of the tokenizer, lightweight decoder, and augmentations from the vanilla ViT backbone; we will expand the discussion in §4 to emphasize these results and to show that the non-hierarchical transformer is essential for the reported performance levels. This will make the source of the gains transparent. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical claim with no derivation chain

full rationale

The paper makes no first-principles derivation or mathematical claim. Its central assertion is that a tokenizer + lightweight decoder + augmentations suffice for a non-hierarchical ViT to reach competitive segmentation performance on nuScenes, SemanticKITTI and Waymo; this is supported solely by empirical comparisons and ablations. No equations, fitted parameters renamed as predictions, self-citations used as uniqueness theorems, or ansatzes smuggled via prior work appear in the provided text. The architecture choices are presented as engineering decisions whose effectiveness is measured externally on public benchmarks, satisfying the criteria for a self-contained, non-circular result.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract introduces no mathematical axioms, free parameters, or new postulated entities; all claims rest on standard supervised training of a transformer on labeled point-cloud datasets.

pith-pipeline@v0.9.1-grok · 5700 in / 1144 out tokens · 22664 ms · 2026-06-28T22:33:33.512961+00:00 · methodology

0 comments
read the original abstract

Plain Transformers have become the de-facto architecture for processing text, audio, image, and video, offering a unified backbone for multimodal learning. However, state-of-the-art architectures for point cloud semantic segmentation remain dominated by U-Nets architectures where convolutions are interleaved with local or windowed attentions. In this work, we show how to effectively leverage vanilla, non-hierarchical ViTs for segmentation of large-scale automotive lidar scenes. We bridge the performance gap thanks to a carefully designed tokenizer, a lightweight decoder segmentation head, and tailored data augmentations. Our approach, VaViT for Vanilla ViT, matches or exceeds the performance of state-of-the-art methods while maintaining the simplicity of ViT architecture. We provide extensive evaluations on nuScenes, SemanticKITTI, and Waymo Open Dataset to validate the efficiency of our method. Code and models are available at https://github.com/valeoai/VaViT.

Figures

Figures reproduced from arXiv: 2605.31177 by Alexandre Boulch, Gilles Puy, Nermin Samet, Renaud Marlet, Spyros Gidaris, Tuan-Hung Vu.

Figure 1
Figure 1. Figure 1: Overview. Our tokenizer aggregates point-level embeddings pi into Q non-empty pillar embeddings, which serves as input tokens tq for a vanilla Vision Transformer. After being processed by the ViT, the Q pillar tokens are redistributed and merged with the original point embeddings, forming the final representations used for point classification. Positions on the 2D BEV space are encoded using RoPE. p (Lemb)… view at source ↗
Figure 2
Figure 2. Figure 2: Attention maps for each head at the 1 st (first) layer of our VaViT-B model trained on nuScenes. The query point, denoted by a red cross, is located on the sidewalk. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Attention maps for each head at the 6 th (middle) layer of our VaViT-B model trained on nuScenes. The query point, denoted by a red cross, is located on the sidewalk. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Attention maps for each head at the 12th (final) layer of our VaViT-B model trained on nuScenes. The query point, denoted by a red cross, is located on the sidewalk. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Attention maps for each head at the 1 st (first) layer of our VaViT-B model trained on nuScenes. The query point, denoted by a red cross, is located on a car. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Attention maps for each head at the 6 th (middle) layer of our VaViT-B model trained on nuScenes. The query point, denoted by a red cross, is located on a car. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Attention maps for each head at the 12th (last) layer of our VaViT-B model trained on nuScenes. The query point, denoted by a red cross, is located on a car. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Attention maps for each head at the 1 st (first) layer of our VaViT-B model trained on SemanticKITTI. The query point, denoted by a red cross, is located on the sidewalk. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p019… view at source ↗
Figure 9
Figure 9. Figure 9: Attention maps for each head at the 6 th (middle) layer of our VaViT-B model trained on SemanticKITTI. The query point, denoted by a red cross, is located on the sidewalk. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p02… view at source ↗
Figure 10
Figure 10. Figure 10: Attention maps for each head at the 12th (last) layer of our VaViT-B model trained on SemanticKITTI. The query point, denoted by a red cross, is located on the sidewalk. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p021… view at source ↗
Figure 11
Figure 11. Figure 11: Attention maps for each head at the 1 st (first) layer of our VaViT-B model trained on SemanticKITTI. The query point, denoted by a red cross, is located on a car. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Attention maps for each head at the 6 th (middle) layer of our VaViT-B model trained on SemanticKITTI. The query point, denoted by a red cross, is located on a car. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Attention maps for each head at the 12th (last) layer of our VaViT-B model trained on SemanticKITTI. The query point, denoted by a red cross, is located on a car. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Attention maps for each head at the 1 st (first) layer of our VaViT-B model trained on WOD. The query point, denoted by a red cross, is located on a car. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Attention maps for each head at the 6 th (middle) layer of our VaViT-B model trained on WOD. The query point, denoted by a red cross, is located on a car. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Attention maps for each head at the 12th (last) layer of our VaViT-B model trained on WOD. The query point, denoted by a red cross, is located on a car. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p027_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Attention maps for each head at the 1 st (first) layer of our VaViT-B model trained on WOD. The query point, denoted by a red cross, is located on a pedestrian. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Attention maps for each head at the 6 th (middle) layer of our VaViT-B model trained on WOD. The query point, denoted by a red cross, is located on a pedestrian. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p029_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Attention maps for each head at the 12th (last) layer of our VaViT-B model trained on WOD. The query point, denoted by a red cross, is located on a pedestrian. Ground truth in BEV is presented at the top. Subsequent maps have a transparency scaled by the attention weight between the query point and the keys; points with zero attention are fully transparent [PITH_FULL_IMAGE:figures/full_fig_p030_19.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

63 extracted references · 3 canonical work pages

  1. [1]

    RangeViT: Towards Vision Transformers for 3D Se- mantic Segmentation in Autonomous Driving

    Angelika Ando, Spyros Gidaris, Andrei Bursuc, Gilles Puy, Alexandre Boulch, and Renaud Marlet. RangeViT: Towards Vision Transformers for 3D Se- mantic Segmentation in Autonomous Driving. In CVPR, 2023. 1, 3

    2023

  2. [2]

    Vivit: A video vision transformer

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lu ˇci´c, and Cordelia Schmid. Vivit: A video vision transformer. InICCV, 2021. 1

    2021

  3. [3]

    Behley, M

    J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. InICCV, 2019. 2, 5, 7, 12

    2019

  4. [4]

    Fkaconv: Feature-kernel alignment for point cloud convolution

    Alexandre Boulch, Gilles Puy, and Renaud Marlet. Fkaconv: Feature-kernel alignment for point cloud convolution. InACCV, 2020. 2

    2020

  5. [5]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krish- nan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krish- nan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driv- ing. InCVPR, 2020. 2, 5, 7, 12

    2020

  6. [6]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herve J ’egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin1. Emerging properties in self-supervised vision transformers. InICCV, 2021. 5

    2021

  7. [7]

    (AF)2-S3Net: Attentive Feature Fusion With Adaptive Feature Selection for Sparse Se- mantic Segmentation Network

    Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (AF)2-S3Net: Attentive Feature Fusion With Adaptive Feature Selection for Sparse Se- mantic Segmentation Network. InCVPR, 2021. 2

    2021

  8. [8]

    PointMixer: MLP- Mixer for Point Cloud Understanding

    Jaesung Choe, Chunghyun Park, Francois Rameau, Jaesik Park, and In So Kweon. PointMixer: MLP- Mixer for Point Cloud Understanding. InECCV,

  9. [9]

    4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks

    Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. InCVPR, 2019. 2

    2019

  10. [10]

    SalsaNext: Fast, Uncertainty-Aware Semantic Segmentation of LiDAR Point Clouds

    Tiago Cortinhal, George Tzelepis, and Eren Erdal Ak- soy. SalsaNext: Fast, Uncertainty-Aware Semantic Segmentation of LiDAR Point Clouds. InAdvances in Visual Computing, 2020. 2

    2020

  11. [11]

    Scaling vision trans- formers to 22 billion parameters

    Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, An- dreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision trans- formers to 22 billion parameters. InICML, 2023. 1

    2023

  12. [12]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InICLR, 2021. 1, 4, 5

    2021

  13. [13]

    TORNADO-Net: mulTiview tOtal vaRiatioN semAntic segmentation with Diamond in- ceptiOn module

    Martin Gerdzhev, Ryan Razani, Ehsan Taghavi, and Liu Bingbing. TORNADO-Net: mulTiview tOtal vaRiatioN semAntic segmentation with Diamond in- ceptiOn module. InICRA, 2021. 2

    2021

  14. [14]

    AST: Audio Spectrogram Transformer

    Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer.arXiv preprint arXiv:2104.01778, 2021. 1

    work page Pith review arXiv 2021

  15. [15]

    Rotary position embedding for vision trans- former

    Byeongho Heo, Song Park, Dongyoon Han, and Sang- doo Yun. Rotary position embedding for vision trans- former. InECCV, 2024. 4

    2024

  16. [16]

    Point-to-V oxel Knowledge Dis- tillation for LiDAR Semantic Segmentation

    Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-V oxel Knowledge Dis- tillation for LiDAR Semantic Segmentation. InCVPR,

  17. [17]

    Deep networks with stochastic depth

    Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth. InECCV, 2016. 7

    2016

  18. [18]

    Dino in the room: Leveraging 2d founda- tion models for 3d segmentation.CVPRW, 2025

    Karim Knaebel, Kadir Yilmaz, Daan de Geus, Alexan- der Hermans, David Adrian, Timm Linder, and Bas- tian Leibe. Dino in the room: Leveraging 2d founda- tion models for 3d segmentation.CVPRW, 2025. 6

    2025

  19. [19]

    KPRNet: Improving projection-based LiDAR semantic segmentation.arXiv:2007.12668,

    Deyvid Kochanov, Fatemeh Karimi Nejadasl, and Olaf Booij. KPRNet: Improving projection-based LiDAR semantic segmentation.arXiv:2007.12668,

    work page arXiv 2007

  20. [20]

    Rethinking range view representation for lidar segmentation

    Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. InICCV, 2023. 1, 3, 6

    2023

  21. [21]

    Lasermix for semi-supervised lidar semantic seg- mentation

    Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic seg- mentation. InCVPR, 2023. 3

    2023

  22. [22]

    Stratified Transformer for 3D Point Cloud Segmenta- tion

    Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Heng- shuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified Transformer for 3D Point Cloud Segmenta- tion. InCVPR, 2022. 2

    2022

  23. [23]

    Spherical Transformer for LiDAR-Based 3D Recognition

    Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical Transformer for LiDAR-Based 3D Recognition. InCVPR, 2023. 2, 6

    2023

  24. [24]

    Large-scale point cloud semantic segmentation with superpoint graphs

    Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. InCVPR, 2018. 2

    2018

  25. [25]

    Self-distillation for robust lidar semantic segmentation in autonomous driving

    Jiale Li, Hang Dai, and Yong Ding. Self-distillation for robust lidar semantic segmentation in autonomous driving. InECCV, 2022. 2, 6

    2022

  26. [26]

    AMVNet: Assertion-based Multi-View Fusion Network for LiDAR Semantic Segmentation

    Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. AMVNet: Assertion-based Multi-View Fusion Network for LiDAR Semantic Segmentation. arXiv:2012.04934, 2020. 2

    work page arXiv 2012

  27. [27]

    Flatformer: Flattened window atten- tion for efficient point cloud transformer

    Zhijian Liu, Xinyu Yang, Haotian Tang, Shang Yang, and Song Han. Flatformer: Flattened window atten- tion for efficient point cloud transformer. InCVPR,

  28. [28]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InICLR, 2019. 7

    2019

  29. [29]

    Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework

    Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework. InICLR, 2022. 2

    2022

  30. [30]

    RangeNet ++: Fast and Accurate Li- DAR Semantic Segmentation

    Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyrill Stachniss. RangeNet ++: Fast and Accurate Li- DAR Semantic Segmentation. InIROS, 2019. 2

    2019

  31. [31]

    Fast Point Transformer

    Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast Point Transformer. InCVPR, 2022. 2

    2022

  32. [32]

    PCSCNet: Fast 3D semantic segmentation of LiDAR point cloud for autonomous car using point convolution and sparse convolution network.Expert Systems with Applications, 2023

    Jaehyun Park, Chansoo Kim, and Kichun Jo Soyeong Kim and. PCSCNet: Fast 3D semantic segmentation of LiDAR point cloud for autonomous car using point convolution and sparse convolution network.Expert Systems with Applications, 2023. 2

    2023

  33. [33]

    Using a waffle iron for automotive point cloud seman- tic segmentation

    Gilles Puy, Alexandre Boulch, and Renaud Marlet. Using a waffle iron for automotive point cloud seman- tic segmentation. InICCV, 2023. 2, 3, 6, 7, 8

    2023

  34. [34]

    Three pillars improving vi- sion foundation model distillation for lidar

    Gilles Puy, Spyros Gidaris, Alexandre Boulch, Oriane Sim´eoni, Corentin Sautier, Patrick P´erez, Andrei Bur- suc, and Renaud Marlet. Three pillars improving vi- sion foundation model distillation for lidar. InCVPR,

  35. [35]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. InCVPR, 2017. 2, 3

    2017

  36. [36]

    PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. InNeurIPS,

  37. [37]

    PointNeXt: Revisit- ing PointNet++ with Improved Training and Scaling Strategies

    Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Abed Al Kader Hammoud, Mohamed Elho- seiny, and Bernard Ghanem. PointNeXt: Revisit- ing PointNet++ with Improved Training and Scaling Strategies. InNeurIPS, 2022. 2

    2022

  38. [38]

    GFNet: Geometric Flow Network for 3D Point Cloud Seman- tic Segmentation.TMLR, 2022

    Haibo Qiu, Baosheng Yu, and Dacheng Tao. GFNet: Geometric Flow Network for 3D Point Cloud Seman- tic Segmentation.TMLR, 2022. 2, 6

    2022

  39. [39]

    Rist, David Schmidt, Markus Enzweiler, and Dariu M

    Christoph B. Rist, David Schmidt, Markus Enzweiler, and Dariu M. Gavrila. SCSSnet: Learning Spatially- Conditioned Scene Segmentation on LiDAR Point Clouds. InIEEE Intelligent Vehicles Symposium,

  40. [40]

    Efficient 3d semantic segmentation with superpoint transformer

    Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. InICCV, 2023. 2

    2023

  41. [41]

    LMSCNet: Lightweight Multiscale 3D Se- mantic Completion

    Luis Rold ˜ao, Raoul de Charette, and Anne Verroust- Blondet. LMSCNet: Lightweight Multiscale 3D Se- mantic Completion. In3DV, 2020. 2

    2020

  42. [42]

    Roformer: Enhanced transformer with rotary position embedding.Neuro- computing, 2023

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neuro- computing, 2023. 4

    2023

  43. [43]

    Scalability in perception for au- tonomous driving: Waymo open dataset

    Pei Sun, , Henrik Kretzschmar, Xerxes Dotiwalla, Aur´elien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vi- jay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Sheng Zhao, Shuyang Cheng, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Ang...

  44. [44]

    Searching effi- cient 3d architectures with sparse point-voxel convo- lution

    Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching effi- cient 3d architectures with sparse point-voxel convo- lution. InECCV, 2020. 2

    2020

  45. [45]

    Qi, Jean-Emmanuel De- schaud, Beatriz Marcotegui, Francois Goulette, and Leonidas J

    Hugues Thomas, Charles R. Qi, Jean-Emmanuel De- schaud, Beatriz Marcotegui, Francois Goulette, and Leonidas J. Guibas. KPConv: Flexible and De- formable Convolution for Point Clouds. InICCV,

  46. [46]

    MLP-Mixer: An all-MLP Architecture for Vision

    Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. MLP-Mixer: An all-MLP Architecture for Vision. InNeurIPS, 2021. 2

    2021

  47. [47]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1

    2017

  48. [48]

    Dy- namic graph cnn for learning on point clouds.ACM Transactions On Graphics, 2019

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dy- namic graph cnn for learning on point clouds.ACM Transactions On Graphics, 2019. 2, 3

    2019

  49. [49]

    Point transformer v3: Simpler faster stronger

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. InCVPR, 2024. 1, 2, 6, 7

    2024

  50. [50]

    Semi-supervised 3d object detec- tion with patchteacher and pillarmix

    Xiaopei Wu, Liang Peng, Liang Xie, Yuenan Hou, Binbin Lin, Xiaoshui Huang, Haifeng Liu, Deng Cai, and Wanli Ouyang. Semi-supervised 3d object detec- tion with patchteacher and pillarmix. InAAAI, 2024. 3, 5, 8

    2024

  51. [51]

    PolarMix: A Gen- eral Data Augmentation Technique for LiDAR Point Clouds

    Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. PolarMix: A Gen- eral Data Augmentation Technique for LiDAR Point Clouds. InNeurIPS, 2022. 3, 8

    2022

  52. [52]

    SqueezeSegV3: Spatially-Adaptive Convolution for Efficient Point-Cloud Segmentation

    Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. SqueezeSegV3: Spatially-Adaptive Convolution for Efficient Point-Cloud Segmentation. InECCV, 2020. 2

    2020

  53. [53]

    RPVNet: A Deep and Efficient Range-Point-V oxel Fusion Network for LiDAR Point Cloud Segmentation

    Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. RPVNet: A Deep and Efficient Range-Point-V oxel Fusion Network for LiDAR Point Cloud Segmentation. InICCV, 2021. 2, 3, 6, 7

    2021

  54. [54]

    Mul- timodal learning with transformers: A survey.IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2023

    Peng Xu, Xiatian Zhu, and David A Clifton. Mul- timodal learning with transformers: A survey.IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2023. 1

    2023

  55. [55]

    2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds

    Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds. InECCV, 2022. 7

    2022

  56. [56]

    Efficient Point Cloud Segmentation with Geometry-Aware Sparse Networks

    Maosheng Ye, Rui Wan, Tongyi Cao Shuangjie Xu, and Qifeng Chen. Efficient Point Cloud Segmentation with Geometry-Aware Sparse Networks. InECCV,

  57. [57]

    Litept: Lighter yet stronger point transformer.CVPR, 2026

    Yuanwen Yue, Damien Robert, Jianyuan Wang, Sunghwan Hong, Jan Dirk Wegner, Christian Rup- precht, and Konrad Schindler. Litept: Lighter yet stronger point transformer.CVPR, 2026. 1, 2, 6, 12

    2026

  58. [58]

    Scaling vision transformers

    Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022. 1

    2022

  59. [59]

    Deep FusionNet for Point Cloud Semantic Seg- mentation

    Feihu Zhang, Jin Fang, Benjamin Wah, and Philip Torr. Deep FusionNet for Point Cloud Semantic Seg- mentation. InECCV, 2020. 2

    2020

  60. [60]

    PolarNet: An Improved Grid Representation for On- line LiDAR Point Clouds Semantic Segmentation

    Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. PolarNet: An Improved Grid Representation for On- line LiDAR Point Clouds Semantic Segmentation. In CVPR, 2020. 2

    2020

  61. [61]

    Torr, and Vladlen Koltun

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, and Vladlen Koltun. Point Transformer. In ICCV, 2021. 2

    2021

  62. [62]

    SV ASeg: Sparse V oxel-Based Attention for 3D LiDAR Point Cloud Semantic Segmentation

    Lin Zhao, Siyuan Xu, Liman Liu, Delie Ming, and Wenbing Tao. SV ASeg: Sparse V oxel-Based Attention for 3D LiDAR Point Cloud Semantic Segmentation. Remote Sens., 2022. 2

    2022

  63. [63]

    Cylindrical and Asymmetrical 3D Convolution Net- works for LiDAR Segmentation

    Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Net- works for LiDAR Segmentation. InCVPR, 2021. 2, 6 A. Appendix A. Details about our adaptation of FlatFormer FlatFormer [27] was designed for object detection using a BEV representation of point clouds. It uses a fla...

    2021