Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

SWA-SOP: Spatially-aware Window Attention for Semantic Occupancy Prediction in Autonomous Driving

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Explicit spatial context inside window attention improves semantic occupancy prediction on both LiDAR and camera inputs, reaching 27.4 mIoU on SemanticKITTI LiDAR and lifting a camera baseline from 12.20 to 13.19 mIoU.

desk verdict A real attention module for LiDAR-based SOP with a solid validation ablation, but the abstract overclaims SOTA and the camera plug-in result is confounded by a full backbone swap. read the letter →

arxiv 2506.18785 v2 pith:EAPE35UZ submitted 2025-06-23 cs.CV cs.AIcs.RO

Helin Cao , Rafael Materla , Sven Behnke This is my paper · ORCID
classification cs.CVcs.AIcs.RO
keywords SemanticoccupancypredictionSpatially-awareattentionSlidingwindowScenecompletionLiDARperceptionCamera-based3DKITTI
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

The paper argues that transformer-based semantic occupancy prediction struggles to complete sparse or occluded 3D scenes because attention lacks explicit spatial structure, and that a sliding-window attention mechanism with per-position spatial embeddings and a center query fixes this. It introduces Spatially-aware Window Attention (SWA), a module that computes attention only inside local windows containing at least one active voxel, using position-aware keys and values plus a query anchored at the window center. On SemanticKITTI's hidden test set, a LiDAR-based pipeline with SWA reaches 57.9 IoU and 27.4 mIoU, ranking among the top results, while swapping deformable attention for SWA inside a camera-based VoxFormer pipeline raises mIoU from 12.20 to 13.19. The significance is that SWA is a plug-in component, so the spatial-context mechanism could transfer to other 3D perception architectures rather than only this one pipeline.

What carries the argument

The carrying object is the Spatially-aware Window Attention (SWA) module, a sliding-window attention that replaces convolution in a 3D U-Net. For each h×w×d window, every voxel slot gets its own feedforward layer, shared across attention heads, that modulates keys and values to encode relative position; a center query is built from the center voxel feature if present or from its global grid index if not; and per-head position weights rescale softmax scores like learnable convolution kernels. Attention runs only when the window contains an active voxel, which keeps cost tied to occupied space and lets the decoder propagate features into empty voxels initialized with shared learnable embeddings.

What would settle it

Evaluate SWA-SOP with per-depth-bin mIoU on SemanticKITTI for voxels beyond 30 m; if the gain over a deformable-attention baseline vanishes or reverses in the farthest bins, the claim that spatial window attention improves completion in sparse distant regions is contradicted.

Watch

Extended reading notes

Core claim

The central claim is that explicitly modeling local spatial context inside attention, rather than relying on global token positions, is what lets a sparse voxel network infer semantics into unobserved regions. In SWA, each window's keys and values receive a position-dependent modulation from slot-specific feedforward layers, and the attention query is taken from the window center, falling back to a learned embedding of the global grid coordinate when the center voxel is empty. The module slides across the volume with stride 2 and computes attention only where at least one voxel is active, so the encoder compresses features over valid regions and the decoder expands them into empty slots that carry shared learnable embeddings. According to the paper, this yields balanced geometry and semantics: state-of-the-art or near-state-of-the-art LiDAR results and consistent gains when the same block is inserted into a camera-based occupancy pipeline.

Load-bearing premise

The mechanism assumes that a local window containing at least one active voxel, plus shared learnable embeddings for the empty slots, provides enough context to infer semantics into unobserved geometry; if large or distant empty regions lack nearby active voxels, the claimed completion gains would not transfer.

Editorial extensions

If this is right

  • A plug-in replacement for deformable or dense attention in existing SOP backbones should reproduce the reported gains without architectural changes outside the attention block.
  • Because windows activate only where voxels are non-empty, compute scales with occupied volume, so larger or higher-resolution occupancy grids become feasible at a fixed budget.
  • The LiDAR pipeline's 57.9 IoU and 27.4 mIoU, together with the camera pipeline's 13.19 mIoU, establish a new joint reference for single-frame completion on SemanticKITTI.
  • The center-query mechanism should keep attention stable in regions where the center voxel is empty, which is exactly where depth-guided sampling tends to fail in image-based methods.

Reading between the lines

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

  • One implicit design choice the paper does not stress is that shared learnable embeddings for empty slots act as a learned prior over unobserved geometry; a testable prediction is that completion quality will be higher for categories with stereotyped shapes, such as roads and buildings, than for rare objects, because the prior has more training examples to encode.
  • The paper attributes camera gains to robustness in far, sparse regions; slicing SemanticKITTI metrics by depth would show whether the improvement concentrates at long range or is distributed across the scene, refining the claimed mechanism.
  • SWA's per-slot feedforward layers over a 3×3×3 window are effectively a learned, data-dependent convolution kernel, suggesting the module could transfer to LiDAR semantic segmentation or detection, where completion is not the goal but local geometric context is.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes Spatially-aware Window Attention (SWA) for semantic occupancy prediction. SWA operates over a voxelized LiDAR feature grid produced by SphereFormer, using sliding windows with conditional activation, per-slot spatial embedding feedforwards, and a center query that falls back to a learned position embedding when the center voxel is empty. The method is evaluated on SemanticKITTI: the LiDAR pipeline reports 57.9 IoU and 27.4 mIoU on the hidden test set, while a camera experiment replacing VoxFormer-S's deformable attention with an adapted SWA U-Net reports 13.19 mIoU and 44.20 IoU. Validation ablations compare the full model, a variant without the center query, and a convolutional U-Net.

Significance. If the central claims hold, SWA would be a useful attention primitive for sparse 3D semantic occupancy prediction, with a coherent sliding-window formulation and a sensible mechanism for propagating information into empty voxels. The paper's strengths include evaluation on the official SemanticKITTI hidden test split, comparison against published baselines, an ablation table isolating the proposed components on validation data, and an attempt to demonstrate cross-modal transfer to a camera-based pipeline. However, the headline 'state-of-the-art' claim is not supported by Table I, and the cross-modal plug-in result is not an isolated test of the attention kernel because it replaces deformable attention with an entire U-Net torso. The central mechanism is internally consistent, and I see no circularity in the evaluation, but the empirical evidence needs strengthening before the paper's claims can be accepted.

major comments (4)
  1. [Abstract; Table I] The abstract and Section V state that the method achieves state-of-the-art results, but Table I shows SSA-SC with higher IoU (58.8 vs 57.9) and S3CNet with higher mIoU (29.5 vs 27.4). The method is second on both headline metrics, as Section IV-B also acknowledges. Please replace the SOTA claim with a qualified statement such as 'competitive' or 'second-best,' or define and compute a joint metric that justifies the claim.
  2. [Section IV-C; Table II] The cross-modality experiment does not isolate SWA. Replacing VoxFormer's deformable attention with 'an adapted SWA U-Net' using proposal queries as input voxels changes the architecture into a full U-Net with down/up-sampling and skip connections, adding many parameters beyond the attention kernel. Since the VoxFormer-S baseline is the published number (12.20 mIoU) and the +SWA model is trained separately, the +0.99 mIoU and +1.25 IoU gains cannot be attributed to SWA alone. A matched control with the same backbone, training budget, and only the attention kernel changed is required to support the plug-in claim.
  3. [Section IV-D; Table III] All comparisons are single runs without error bars. The center-query ablation gap (27.91 vs 27.07 mIoU) is small, and the full-model versus convolutional-U-Net comparison (27.91 vs 24.46) changes architecture and capacity, not only the attention mechanism. Repeated seeds and matched parameter/computation counts are needed to establish that the reported gains are significant and attributable to the proposed components.
  4. [Section III-A; Table II] The claimed scene-completion ability rests on the assumption that shared learnable embeddings for empty voxels, combined with a center query, can propagate semantics into unobserved regions. The paper does not report any distance- or density-stratified analysis of completion, and the only cross-modality experiment does not isolate the completion mechanism. Please add an evaluation of completion quality (for example, IoU as a function of range or voxel-density buckets) or an ablation in which the decoder expansion into empty regions is disabled, so that the geometric-expansion claim is directly tested.
minor comments (6)
  1. [Introduction] References [1]–[3] are cited as transformer-based SOP methods, but [1] is a facial-expression recognition paper, [2] a video super-resolution paper, and [3] a medical image segmentation paper; these citations do not support the sentence they are attached to.
  2. [Eq. (3)] The notation is inconsistent: 'if the fc is non-empty' should be 'if fc is non-empty', and it is unclear how the 3D integer index pc is embedded before being passed to FFq; please specify the embedding layer explicitly.
  3. [Section IV-A] Implementation details omit batch size, the number of windows or model parameters, inference time, and the exact treatment of empty versus unknown voxels during training; please add these details to support reproducibility.
  4. [Table I] The table header 'SC SSC' is confusing, and the per-class columns are not labeled as percentages or IoU values; the method order is also unsorted, which makes the table harder to read.
  5. [Section IV-B] The discussion about S3CNet benefiting from additional 2D BEV features is not backed by any quantitative comparison in the paper and reads as post-hoc justification; either add an experiment or soften the claim.
  6. [Figure 5] The caption says all 19 semantic classes are rendered, but the figure has no color legend, making the qualitative comparison difficult to interpret; please add a legend.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: SWA is defined from trainable components and benchmarked against the external SemanticKITTI test set; self-citations are background only.

full rationale

The derivation chain is self-contained, and no prediction reduces to its inputs by construction. The SWA mechanism (Section III, Eqs. 1-5) is fully defined from trainable layers (FF projections, per-slot modulation, softmax reweighting) and voxel validity; none of these quantities is defined in terms of the evaluation metrics (IoU, mIoU), and the paper never fits a parameter to the hidden test set. All headline numbers (Tables I-III) are produced by training on the official SemanticKITTI split (sequences 00-10 excluding 08) and evaluating on the validation or hidden test set, so the 'fitted input called prediction' pattern does not apply. The self-citations ([14] SLCF-Net, [21] OC-SOP, [25] DiffSSC) appear only as related-work background and are not load-bearing: they supply no theorem, no fitted value, and no argument forbidding alternatives, so they do not constitute self-citation circularity; the SemanticKITTI dataset citation is the external benchmark. The camera-based plug-in experiment (Section IV-C, Table II) replaces VoxFormer's deformable attention with 'an adapted SWA U-Net,' which is a confound for attributing the +0.99 mIoU gain to the attention kernel, but the reported 13.19 mIoU is an independently measured result rather than one constructed from its inputs, so this is an experimental-isolation concern for correctness risk, not circularity. Two manuscript-internal issues are flagged without affecting the circularity score: the abstract's 'state-of-the-art results' claim is contradicted by Table I and by the paper's own Section IV-B, where SWA-SOP ranks second on both IoU (57.9 vs SSA-SC 58.8) and mIoU (27.4 vs S3CNet 29.5); and references [1]-[3], cited for 'Transformer-based architectures' in SOP, concern facial-expression recognition, video super-resolution, and medical image segmentation, i.e., missing support for that sentence. Neither issue makes any equation or evaluation equivalent to its input. Self-containment against the external benchmark and the validation ablation (A vs C: 27.91 vs 24.46 mIoU) supports the honest non-finding of circularity.

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

The method introduces learned components (spatial FFNs, modulation weights, center-query MLP, empty-voxel embeddings) and many hand-set hyperparameters; none are derived from first principles, and all are fit to training data. This is normal for supervised deep learning, but it means the reported result is a trained model output, not a parameter-free derivation.

free parameters (5)
  • Attention window size, stride, padding = 3x3x3 window, stride 2, padding 1
    Chosen by hand; defines the locality of attention and the density of the sliding traversal; no derivation is given.
  • Number of heads and hidden dimension = 8 heads, d=128
    Chosen architecture hyperparameters; not derived from data or theory.
  • Training length and learning rate schedule = 50 epochs, initial LR 0.006, polynomial decay power 0.9
    Standard schedule choices; affects results but is not part of the method's derivation.
  • Per-slot spatial embedding FFNs and per-head modulation weights = Learned during training
    The core mechanism's parameters are fit to the training set; there is no analytic or external constraint.
  • Shared learnable embeddings for empty voxels = Learned during training
    Used to initialize decoder positions so attention can densify the volume.
assumptions (5)
  • standard math Dot-product attention with softmax and head-specific projections is an appropriate aggregation mechanism for 3D voxel features.
    Used throughout Equations (1) to (5) without proof; standard attention machinery.
  • domain assumption SemanticKITTI voxelized cumulative scans and the unknown mask provide valid ground truth for occupancy and semantics.
    Dataset convention inherited from [18]; dynamic objects accumulated over time can create artifacts, as the paper itself notes in Section IV-B.
  • ad hoc to paper Providing a valid query from the center voxel, or from global grid coordinates when the center is empty, supplies enough context for attention in sparse regions.
    Center query design in Section III-C; its effectiveness is shown only by ablation, not by an independent argument.
  • ad hoc to paper Initializing empty voxels with shared learnable embeddings and propagating features into them yields meaningful scene completion.
    Decoder expansion assumption in Section III-A; no external evidence beyond the reported benchmark gains.
  • domain assumption Point-wise semantic logits from SphereFormer, concatenated with intermediate features, form a sufficient sparse voxel representation for the SOP U-Net.
    Pipeline design in Section III; the choice of SphereFormer as fixed feature extractor is not ablated.
invented entities (3)
  • Synthetic center query from global grid index
    purpose: Produces a valid query when the center voxel is empty, so attention can still aggregate window context.
    A new mechanism introduced in Section III-C; its benefit is supported only by the paper's own ablation.
  • Per-slot spatial embedding feedforward layers
    purpose: Modulate keys and values by slot position to inject local spatial geometry into attention.
    Introduced in Section III-B; no external validation outside this paper.
  • Shared learnable embeddings for empty decoder voxels
    purpose: Provide features for empty voxels so attention can densify unobserved regions.
    Introduced in Section III-A; not independently evidenced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SWA-SOP: Spatially-aware Window Attention for Semantic Occupancy Prediction in Autonomous Driving." pith.science (2026). https://pith.science/paper/EAPE35UZ

@misc{pith2026250618785,
  author       = {Pith},
  title        = {Pith review of: SWA-SOP: Spatially-aware Window Attention for Semantic Occupancy Prediction in Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EAPE35UZ}},
  note         = {Machine review of arXiv:2506.18785}
}
read the original abstract

Perception systems in autonomous driving rely on sensors such as LiDAR and cameras to perceive the 3D environment. However, due to occlusions and data sparsity, these sensors often fail to capture complete information. Semantic Occupancy Prediction (SOP) addresses this challenge by inferring both occupancy and semantics of unobserved regions. Existing transformer-based SOP methods lack explicit modeling of spatial structure in attention computation, resulting in limited geometric awareness and poor performance in sparse or occluded areas. To this end, we propose Spatially-aware Window Attention (SWA), a novel mechanism that incorporates local spatial context into attention. SWA significantly improves scene completion and achieves state-of-the-art results on LiDAR-based SOP benchmarks. We further validate its generality by integrating SWA into a camera-based SOP pipeline, where it also yields consistent gains across modalities.

Figures

Figures reproduced from arXiv: 2506.18785 by the authors.

Figure 1
Figure 1. SWA-SOP estimates a dense, semantically labeled scene (b) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall pipeline of SWA-SOP. The raw LiDAR point cloud is first semantically segmented by SphereFormer [30] and then voxelized [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the spatial embedding module. All valid [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Overview of the intra- and inter-window attention pipeline. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results on the SemanticKITTI validation set. We show the voxelized LiDAR input (RGB image included for visualization [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. OC-SOP: Enhancing Vision-Based 3D Semantic Occupancy Prediction by Object-Centric Awareness

    cs.CV 2025-06 conditional novelty 6.0 of 10

    OC-SOP fuses object detection queries into a semantic occupancy completion U-Net, improving foreground-object voxel accuracy and achieving state-of-the-art mIoU on SemanticKITTI.

Reference graph

Works this paper leans on

42 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [1]

    Cf-dan: Facial-expression recognition based on cross-fusion dual-attention network,

    F. Zhang, G. Chen, H. Wang, and C. Zhang, “Cf-dan: Facial-expression recognition based on cross-fusion dual-attention network,” Computa- tional Visual Media , 2024

    work page 2024

  2. [2]

    Multi-scale video super-resolution transformer with polynomial approximation,

    F. Zhang, G. Chen, H. Wang, J. Li, and C. Zhang, “Multi-scale video super-resolution transformer with polynomial approximation,” IEEE Transactions on Circuits and Systems for Video Technology , 2023

    work page 2023

  3. [3]

    A medical image segmentation model with auto-dynamic convolution and location attention mecha- nism,

    Y . Wang, H. Wang, and F. Zhang, “A medical image segmentation model with auto-dynamic convolution and location attention mecha- nism,” Computer Methods and Programs in Biomedicine , 2025

    work page 2025

  4. [4]

    V oxelNet: End-to-end learning for point cloud based 3D object detection,

    Y . Zhou and O. Tuzel, “V oxelNet: End-to-end learning for point cloud based 3D object detection,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2018

    work page 2018

  5. [5]

    Cylindrical and asymmetrical 3D convolution networks for LiDAR segmentation,

    X. Zhu, H. Zhou, T. Wang, F. Hong, Y . Ma, W. Li, H. Li, and D. Lin, “Cylindrical and asymmetrical 3D convolution networks for LiDAR segmentation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 9939–9948

    work page 2021

  6. [6]

    PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017

    work page 2017

  7. [7]

    PointNet++: Deep hierarchical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “PointNet++: Deep hierarchical feature learning on point sets in a metric space,” Advances in Neural Information Processing Systems (NeurIPS) , vol. 30, 2017

    work page 2017

  8. [8]

    SRKD: Towards efficient 3D point cloud segmentation via Structure- and Relation- aware knowledge distillation,

    Y . Li, J. Dong, Z. Dong, C. Yang, Z. An, and Y . Xu, “SRKD: Towards efficient 3D point cloud segmentation via Structure- and Relation- aware knowledge distillation,” arXiv preprint arXiv:2506.17290, 2025

    arXiv 2025

Show all 42 references
  1. [9]

    Adaptive fusion of lidar features for 3d object detection in autonomous driving,

    M. Wang, D. Li, J. R. Casas, and J. Ruiz-Hidalgo, “Adaptive fusion of lidar features for 3d object detection in autonomous driving,” Sensors, 2025

    2025

  2. [10]

    Center-based 3D object detection and tracking,

    T. Yin, X. Zhou, and P. Krahenbuhl, “Center-based 3D object detection and tracking,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  3. [11]

    MobileStereoNet: Towards lightweight deep networks for stereo matching,

    F. Shamsafar, S. Woerz, R. Rahim, and A. Zell, “MobileStereoNet: Towards lightweight deep networks for stereo matching,” inIEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022

    2022

  4. [12]

    Pseudo-LiDAR from visual depth estimation: Bridg- ing the gap in 3D object detection for autonomous driving,

    Y . Wang, W.-L. Chao, D. Garg, B. Hariharan, M. Campbell, and K. Weinberger, “Pseudo-LiDAR from visual depth estimation: Bridg- ing the gap in 3D object detection for autonomous driving,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2019

    2019

  5. [13]

    Pseudo-LiDAR++: Accurate depth for 3D object detection in autonomous driving,

    Y . You, Y . Wang, W.-L. Chao, D. Garg, G. Pleiss, B. Hariharan, M. Campbell, and K. Q. Weinberger, “Pseudo-LiDAR++: Accurate depth for 3D object detection in autonomous driving,” International Conference on Learning Representations, (ICLR) , 2020

    2020

  6. [14]

    SLCF-Net: Sequential LiDAR-camera fusion for semantic scene completion using a 3D recurrent U-Net,

    H. Cao and S. Behnke, “SLCF-Net: Sequential LiDAR-camera fusion for semantic scene completion using a 3D recurrent U-Net,” in IEEE International Conference on Robotics and Automation (ICRA) , 2024, pp. 2767–2773

    2024

  7. [15]

    Source-free domain adaptation with class prototype discovery,

    L. Zhou, N. Li, M. Ye, X. Zhu, and S. Tang, “Source-free domain adaptation with class prototype discovery,” Pattern Recognition, 2024

    2024

  8. [16]

    Self- prompting analogical reasoning for uav object detection,

    N. Li, M. Ye, L. Zhou, S. Tang, Y . Gan, Z. Liang, and X. Zhu, “Self- prompting analogical reasoning for uav object detection,” in National Conference on Artificial Intelligence (AAAI) , 2025

    2025

  9. [17]

    Semantic scene completion from a single depth image,

    S. Song, F. Yu, A. Zeng, A. X. Chang, M. Savva, and T. Funkhouser, “Semantic scene completion from a single depth image,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017, pp. 1746–1754

    2017

  10. [18]

    SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences

    J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stach- niss, and J. Gall, “SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences.”

  11. [19]

    LMSCNet: Lightweight multiscale 3D semantic completion,

    L. Rold ˜ao, R. de Charette, and A. Verroust-Blondet, “LMSCNet: Lightweight multiscale 3D semantic completion,” in International Conference on 3D Vision (3DV) , 2020

    2020

  12. [20]

    MonoScene: Monocular 3D semantic scene completion,

    A.-Q. Cao and R. De Charette, “MonoScene: Monocular 3D semantic scene completion,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  13. [21]

    OC-SOP: Enhancing Vision-Based 3d seman- tic occupancy prediction by Object-Centric awareness,

    H. Cao and S. Behnke, “OC-SOP: Enhancing Vision-Based 3d seman- tic occupancy prediction by Object-Centric awareness,” arXiv preprint arXiv:2506.18798, 2025

    2025 arXiv

  14. [22]

    V oxFormer: Sparse voxel transformer for camera- based 3d semantic scene completion,

    Y . Li, Z. Yu, C. Choy, C. Xiao, J. M. Alvarez, S. Fidler, C. Feng, and A. Anandkumar, “V oxFormer: Sparse voxel transformer for camera- based 3d semantic scene completion,” in IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) , 2023

    2023

  15. [23]

    OccFormer: Dual-path transformer for vision-based 3d semantic occupancy prediction,

    Y . Zhang, Z. Zhu, and D. Du, “OccFormer: Dual-path transformer for vision-based 3d semantic occupancy prediction,” inIEEE International Conference on Computer Vision (ICCV) , 2023

    2023

  16. [24]

    Deformable DETR: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable DETR: Deformable transformers for end-to-end object detection,” International Conference on Learning Representations, (ICLR) , 2021

    2021

  17. [25]

    DiffSSC: Semantic LiDAR scan comple- tion using denoising diffusion probabilistic models,

    H. Cao and S. Behnke, “DiffSSC: Semantic LiDAR scan comple- tion using denoising diffusion probabilistic models,” arXiv preprint arXiv:2409.18092, 2024

    2024 arXiv

  18. [26]

    Sparse 3D convolutional neural networks,

    B. Graham, “Sparse 3D convolutional neural networks,” in British Machine Vision Conference (BMVC) , 2015

    2015

  19. [27]

    3D semantic segmentation with submanifold sparse convolutional networks,

    B. Graham, M. Engelcke, and L. van der Maaten, “3D semantic segmentation with submanifold sparse convolutional networks,” IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2018

    2018

  20. [28]

    Sparse convolutional neural networks,

    B. Liu, M. Wang, H. Foroosh, M. Tappen, and M. Pensky, “Sparse convolutional neural networks,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2015

    2015

  21. [29]

    Stratified transformer for 3D point cloud segmentation,

    X. Lai, J. Liu, L. Jiang, L. Wang, H. Zhao, S. Liu, X. Qi, and J. Jia, “Stratified transformer for 3D point cloud segmentation,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2022

    2022

  22. [30]

    Spherical transformer for LiDAR-based 3D recognition,

    X. Lai, Y . Chen, F. Lu, J. Liu, and J. Jia, “Spherical transformer for LiDAR-based 3D recognition,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2023

    2023

  23. [31]

    V oxel transformer for 3D object detection,

    J. Mao, Y . Xue, M. Niu et al. , “V oxel transformer for 3D object detection,” in IEEE International Conference on Computer Vision (ICCV), 2021

    2021

  24. [32]

    Semantic scene completion using local deep implicit functions on LiDAR data,

    C. B. Rist, D. Emmerichs, M. Enzweiler, and D. M. Gavrila, “Semantic scene completion using local deep implicit functions on LiDAR data,” IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2021

    2021

  25. [33]

    Semantic segmentation-assisted scene completion for LiDAR point clouds,

    X. Yang, H. Zou, X. Kong, T. Huang, Y . Liu, W. Li, F. Wen, and H. Zhang, “Semantic segmentation-assisted scene completion for LiDAR point clouds,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2021

    2021

  26. [34]

    Sparse single sweep LiDAR point cloud segmentation via learning contextual shape priors from scene completion,

    X. Yan, J. Gao, J. Li, R. Zhang, Z. Li, R. Huang, and S. Cui, “Sparse single sweep LiDAR point cloud segmentation via learning contextual shape priors from scene completion,” in National Conference on Artificial Intelligence (AAAI) , 2021

    2021

  27. [35]

    S3CNnet: A sparse semantic scene completion network for LiDAR point clouds,

    R. Cheng, C. Agia, Y . Ren, X. Li, and L. Bingbing, “S3CNnet: A sparse semantic scene completion network for LiDAR point clouds,” in Proceedings of Machine Learning Research (PMLR) , 2021

    2021

  28. [36]

    Are we ready for autonomous driving? the KITTI vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the KITTI vision benchmark suite,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2012

    2012

  29. [37]

    SpConv: Spatially sparse convolution library,

    SpConv Contributors, “SpConv: Spatially sparse convolution library,” https://github.com/traveller59/spconv, 2022

    2022

  30. [38]

    LLM-Enhanced multimodal fusion for Cross-Domain sequential rec- ommendation,

    W. Wu, Z. Chen, X. Qiu, S. Song, X. Huang, F. Ma, and J. Xiao, “LLM-Enhanced multimodal fusion for Cross-Domain sequential rec- ommendation,” arXiv preprint arXiv:2506.17966 , 2025

    2025

  31. [39]

    Image fusion for cross-domain sequential recommendation,

    W. Wu, S. Song, X. Qiu, X. Huang, F. Ma, and J. Xiao, “Image fusion for cross-domain sequential recommendation,” in ACM Web Conference, 2025

    2025

  32. [40]

    Prompt categories cluster for weakly supervised semantic segmenta- tion,

    W. Wu, X. Qiu, S. Song, Z. Chen, X. Huang, F. Ma, and J. Xiao, “Prompt categories cluster for weakly supervised semantic segmenta- tion,” in IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2025, pp. 3198–3207

    2025

  33. [41]

    SGLP: A similarity guided fast layer partition pruning for compressing large deep models,

    Y . Li, Y . Lu, Z. Dong, C. Yang, Y . Chen, and J. Gou, “SGLP: A similarity guided fast layer partition pruning for compressing large deep models,” arXiv preprint arXiv:2410.14720 , 2024

    2024

  34. [42]

    Frequency-Aligned knowledge distillation for lightweight spatiotemporal forecasting,

    Y . Li, C. Yang, H. Zeng, Z. Dong, Z. An, Y . Xu, Y . Tian, and H. Wu, “Frequency-Aligned knowledge distillation for lightweight spatiotemporal forecasting,” arXiv:2507.02939, 2025

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.