Pith. sign in

REVIEW 3 major objections 5 minor 46 references

DepthMatch: Semi-Supervised RGB-D Scene Parsing through Depth-Guided Regularization

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DepthMatch pushes semi-supervised RGB-D scene parsing to 61.4% mIoU on NYUv2 and first place on the KITTI Semantics benchmark, using three depth-guided components over a UniMatchV2 baseline.

desk verdict A clean, modest semi-supervised RGB-D parsing paper with credible NYUv2 ablations; don't buy the KITTI rank-1 claim until the submission protocol is disclosed. read the letter →

arxiv 2505.20041 v1 pith:ABGPE7JV submitted 2025-05-26 cs.CV

Jianxin Huang , Jiahang Li , Sergey Vityazev , Alexander Dvorkovich , Rui Fan This is my paper
classification cs.CV
keywords semi-supervisedlearningRGB-Dsceneparsingconsistencyregularizationdepth-guidedfeaturefusionboundarylossNYUv2KITTISemantics
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

This paper claims that semi-supervised RGB-D scene parsing can be made both more accurate and more parameter-efficient by leveraging depth information in three targeted ways. The authors argue that existing RGB-D parsers rely on large amounts of pixel-level labels, and their DepthMatch framework reduces this dependence by regularizing unlabeled RGB-D pairs through a complementary patch mix-up, a lightweight depth-to-RGB feature injector, and a depth-guided boundary loss. On NYUv2, DepthMatch reaches 61.4% mean IoU when extra unlabeled SUN-RGBD data is used, surpassing the previous best method; on KITTI it reports 76.60% class mIoU and 90.16% category mIoU, ranking first on the official benchmark. If these results hold, the framework provides a practical recipe for indoor and outdoor scene parsing with limited annotations.

What carries the argument

The framework rests on three mechanisms: CPMA, which replaces patches of the unlabeled RGB image with the corresponding depth patches using complementary random masks, forcing the model to rely on geometric cues; LSPI, which reduces the depth and RGB embeddings to a quarter of their dimensionality, fuses them, and adds the result back to the RGB embedding, injecting spatial priors with minimal parameters; and DGBL, which binarizes boundary maps extracted from the predicted segmentation and from the depth image, applies a mean-squared-error loss weighted by predicted boundary pixels, and decays its weight over training. The teacher-student EMA setup with weak-to-strong consistency provides the learning signal for unlabeled data, using a 0.95 confidence threshold for pseudo-labels.

What would settle it

Run DepthMatch on the KITTI Semantics benchmark with single-scale inference and the same training schedule as the NYUv2 ablations; if the class mIoU drops below the second-ranked method or the gap shrinks to noise, the rank-1 result would come from inference-time scaling rather than the proposed components.

Watch

Extended reading notes

Core claim

DepthMatch establishes that a semi-supervised RGB-D parser can outperform fully supervised multimodal parsers by injecting depth as a spatial prior during encoding, mixing complementary patches between the RGB and depth modalities as a strong augmentation, and supervising predicted boundaries with depth-derived contours on unlabeled data. The paper's central claim is that each of the three components gives a measurable improvement over the UniMatchV2 baseline on NYUv2, with the full system adding 1.8 mIoU over the RGB-only baseline, and that the same framework transfers to outdoor driving scenes where it tops the KITTI Semantics leaderboard with 76.60% class mIoU and 90.16% category mIoU.

Load-bearing premise

The reported KITTI rank and some NYUv2 gains depend on the evaluation protocol matching the ablations, including the multi-scale inference used only for KITTI and the unspecified training details of the official submission.

Editorial extensions

If this is right

  • If DepthMatch's claims are correct, semi-supervised RGB-D parsers can be built with a fraction of the parameters of prior fusion modules while beating supervised baselines on indoor benchmarks.
  • Depth-derived boundary supervision offers a way to impose geometric structure on unlabeled data without ground-truth masks, which could generalize to other dense prediction tasks.
  • The framework's use of ViTAStereo-estimated depth on KITTI suggests the method tolerates imperfect, estimated depth rather than requiring sensor depth.
  • The 1/4-labeled-data run (56.6% mIoU) implies that the method remains competitive under severe annotation scarcity, widening the practical applicability of RGB-D parsing.

Reading between the lines

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

  • The reported KITTI rank is only partly attributable to the proposed modules: because multi-scale inference with scaling factors {1.05, 1.5, 2.0, 2.5} is used only for KITTI, a direct single-scale comparison would be needed to separate the modules' contribution from inference-time scaling.
  • Since the depth signal for KITTI comes from a stereo matcher rather than a real depth sensor, the outdoor results speak to robustness against noisy estimated depth, but the exact depth quality used in the submission is not reported.
  • The complementary patch-swapping idea in CPMA could transfer to other paired modalities such as RGB-thermal or RGB-normal data, where one modality is sparse or noisy and the other carries complementary structure.
  • A testable extension would be ablating DGBL with different boundary extraction thresholds or using depth edges from a stronger depth model, to see whether the boundary supervision depends on the specific contour estimator.
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

3 major / 5 minor

Summary. DepthMatch proposes a semi-supervised RGB-D scene parsing framework built on a mean-teacher consistency-regularization pipeline with three new components: Complementary Patch Mix-up Augmentation (CPMA), which swaps patches between the RGB and depth images for unlabeled data; a Lightweight Spatial Prior Injector (LSPI), which fuses depth features into the RGB embedding stream with linear layers; and a Depth-Guided Boundary Loss (DGBL), which uses depth-derived boundaries to supervise predicted boundaries on unlabeled images. Experiments on NYUv2 show consistent incremental gains in ablation (from 59.6 to 61.4 mIoU) and a reported state-of-the-art result of 61.4% mIoU using SUN-RGBD as unlabeled data. On KITTI Semantics, the paper reports ranking first with 76.60% class mIoU and 90.16% category mIoU, using depth maps produced by ViTAStereo and multi-scale inference.

Significance. If the reported results hold, the paper demonstrates that three relatively simple and lightweight modifications can yield meaningful gains in semi-supervised RGB-D scene parsing over a strong UniMatchV2 baseline. The NYUv2 ablations are internally consistent: each component contributes a positive increment (LSPI +0.6, CPMA +0.7, DGBL +0.6 in the relevant configurations), and the full model reaches 61.4 mIoU with fewer parameters than most compared supervised methods. The authors promise public code, which would aid reproducibility. However, the headline 'rank first on KITTI' claim currently rests on an insufficiently documented protocol, and the DGBL loss has an asymmetric formulation that is not fully justified. These are load-bearing issues that require additional experiments or clear analysis.

major comments (3)
  1. [Section III.A, Table I] The rows 'DepthMatch (1/4)' and 'DepthMatch (1/2)' are not described: the paper never states what fraction of the 795 labeled NYUv2 training images is used, nor whether SUN-RGBD unlabeled data is included in those rows. The text in Section III.A says the dataset is divided into 795 training and 654 testing samples, but the 1/4 and 1/2 splits are never defined. In addition, Section III.D states that all ablation experiments use the SUN-RGBD training set as unlabeled data, while Table I's no-star DepthMatch (59.9 mIoU) appears to correspond to a model trained without unlabeled data, because the starred entry (61.4) is explicitly marked as using SUN-RGBD. This inconsistency makes it impossible to separate the contribution of unlabeled data from the contribution of the proposed modules, and it directly affects the interpretation of the semi-supervised gains claimed in the paper. Please specify the labeled/unlabeled configuration of every row in Table I and reconcile it with the ablation setup.
  2. [Section III.B and III.C, Table II] The claim of ranking first on the KITTI Semantics benchmark is not supported by controlled experiments within the manuscript. Section III.B discloses that multi-scale inference with scaling factors {1.05, 1.5, 2.0, 2.5} is used only for KITTI, but the paper reports no single-scale result, no number of training epochs, no labeled/unlabeled split for the official submission, and no statement that the submitted model used the same hyperparameters as the NYUv2 ablations. Because the leaderboard comparisons in Table II include methods that do not use depth and likely do not use multi-scale inference, the rank-1 result cannot be attributed to CPMA, LSPI, or DGBL; the additional ViTAStereo depth signal and test-time scaling are confounded variables. Please provide a single-scale baseline, an ablation with and without ViTAStereo depth, and report the submission identifiers and retrieval date for the leaderboard entries in Table II.
  3. [Section II.D, Eq. (5)] The DGBL loss is asymmetric in a way that may undermine its stated purpose. The loss is defined as (1/N) * sum_j yp_ij * (yp_ij - yd_ij)^2, so pixels where yp_ij = 0 (i.e., not predicted as boundary) contribute zero loss even if the depth map indicates a boundary there. Consequently, the loss does not penalize missing boundaries, and it also does not penalize predicted boundaries that do not coincide with depth boundaries unless yp_ij = 1 at that pixel. The paper justifies this by saying the loss is only calculated for predicted boundary pixels, but it does not explain why this asymmetric weighting is preferable to a symmetric boundary loss such as (yp_ij - yd_ij)^2. Since DGBL is one of the three core contributions, the formulation should be justified theoretically or compared against symmetric variants in the ablation study.
minor comments (5)
  1. [Section II.C, Eq. (4)] The text says the heterogeneous embeddings are 'downsampled' and then 'upsampled' to their original dimensions, but Eq. (4) uses linear layers that reduce and restore only the feature-channel dimension, not the spatial resolution. Please clarify the terminology to avoid confusion between spatial downsampling and channel reduction.
  2. [Section II.D] The phrase 'ground truth boundaries' for unlabeled data is imprecise; these are depth-derived boundary pseudo-labels, not manual annotations. Please adjust the wording to reflect the pseudo-label nature of yd_ij.
  3. [Equation (2)] The notation pw_i and ˆpw_i is introduced without a clear definition of the argmax operation and the temperature or sharpening step. Please define the pseudo-label generation procedure more explicitly.
  4. [Section III.B] The KITTI depth data is obtained from ViTAStereo, a learned stereo matcher, but the paper does not describe how the stereo disparity is converted to depth or whether any filtering is applied. Please provide this detail for reproducibility.
  5. [Table II] Table II lists only five methods without submission identifiers or dates; leaderboard rankings can change over time. Please include the retrieval date and submission identifiers to allow independent verification.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation chain found; the load-bearing concerns are experimental validity issues, not circularity.

full rationale

DepthMatch's claimed derivation is self-contained and does not reduce any prediction to its inputs by construction. The supervised loss (Eq. 1) uses ground-truth masks, while the unlabeled loss (Eq. 2) uses teacher-generated pseudo-labels filtered by a confidence threshold; this is standard self-training and not a circular fit. CPMA (Eq. 3) is an augmentation that mixes RGB and depth patches, LSPI (Eq. 4) is an architectural fusion module, and DGBL (Eq. 5) compares predicted boundaries against boundaries extracted from the depth image, which is an independent signal. The NyuV2 ablations (Tables III-IV) incrementally add components to a UniMatchV2 baseline, providing internal evidence for each contribution. The self-citations ([8], [12], [13]) are contextual prior-work citations and are not load-bearing. Two validity risks should be weighed separately from circularity: the KITTI rank-1 claim relies on leaderboard comparisons with no single-scale baseline, no stated labeled/unlabeled protocol for the submitted model, and no ablation isolating the contribution of ViTAStereo depth or multi-scale inference; and the use of the SUN-RGBD training set as unlabeled data for NYUv2 does not state whether that split excludes NYUv2 test images. DGBL also only computes loss where the predicted boundary is already positive, so it may under-supervise missing boundaries. These are correctness and verifiability concerns, not instances where the derivation is equivalent to its own input, so the circularity score is low.

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

No new physical or conceptual entities are introduced. The framework relies on standard semi-supervised learning assumptions, pretrained backbones, and depth-derived boundary maps. Free parameters are limited to a few hand-set hyperparameters, of which only the CPMA masking ratio is swept; the value of lambda_u is not even reported.

free parameters (4)
  • confidence threshold tau = 0.95
    Pseudo-label confidence threshold in Eq. (2), chosen by hand following FixMatch/UniMatch; not tuned per dataset and its sensitivity is not reported.
  • CPMA masking ratio = 0.1
    Swept in Table III over {0, 0.05, 0.1, 0.15, 0.2}; 0.1 chosen by best ablation mIoU (60.9% on the NYUv2 SUN-RGBD setting).
  • DGBL binarization threshold = 0.1
    Threshold applied to boundary maps before computing Eq. (5); no sensitivity analysis is provided.
  • lambda_u (unlabeled loss weight) = not stated
    Weight balancing supervised and unsupervised losses in Eq. (6); the value is never reported in the paper.
assumptions (4)
  • domain assumption DINOv2-S self-supervised features generalize across indoor and outdoor RGB-D data.
    The pipeline uses DINOv2-S as fixed encoder features for both NYUv2 and KITTI; if the pretrained features do not transfer, the reported gains would not hold. This is a borrowed pretraining asset, not verified in the paper.
  • domain assumption EMA teacher pseudo-labels are a stable and reliable training signal.
    The student is supervised by teacher-generated pseudo-labels (Section II-A). The paper relies on the standard semi-supervised assumption without independent verification.
  • domain assumption Depth-derived boundaries coincide with semantic object boundaries.
    DGBL (Section II-D) assumes that edges in the depth image provide useful boundary supervision for semantic segmentation. This is not always true, and the paper does not analyze failure cases.
  • domain assumption ViTAStereo depth estimates for KITTI are sufficiently accurate to use as depth input.
    For KITTI, depth is obtained using ViTAStereo [36] rather than a real sensor. The error statistics of this estimated depth are not reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DepthMatch: Semi-Supervised RGB-D Scene Parsing through Depth-Guided Regularization." pith.science (2026). https://pith.science/paper/ABGPE7JV

@misc{pith2026250520041,
  author       = {Pith},
  title        = {Pith review of: DepthMatch: Semi-Supervised RGB-D Scene Parsing through Depth-Guided Regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ABGPE7JV}},
  note         = {Machine review of arXiv:2505.20041}
}
read the original abstract

RGB-D scene parsing methods effectively capture both semantic and geometric features of the environment, demonstrating great potential under challenging conditions such as extreme weather and low lighting. However, existing RGB-D scene parsing methods predominantly rely on supervised training strategies, which require a large amount of manually annotated pixel-level labels that are both time-consuming and costly. To overcome these limitations, we introduce DepthMatch, a semi-supervised learning framework that is specifically designed for RGB-D scene parsing. To make full use of unlabeled data, we propose complementary patch mix-up augmentation to explore the latent relationships between texture and spatial features in RGB-D image pairs. We also design a lightweight spatial prior injector to replace traditional complex fusion modules, improving the efficiency of heterogeneous feature fusion. Furthermore, we introduce depth-guided boundary loss to enhance the model's boundary prediction capabilities. Experimental results demonstrate that DepthMatch exhibits high applicability in both indoor and outdoor scenes, achieving state-of-the-art results on the NYUv2 dataset and ranking first on the KITTI Semantics benchmark.

Figures

Figures reproduced from arXiv: 2505.20041 by the authors.

Figure 1
Figure 1. An overview of our proposed DepthMatch. more than depth images in RGB-D scene parsing tasks, we propose a Lightweight Spatial Prior Injector (LSPI) to replace traditional complex fusion modules, achieving efficient integration of heterogeneous features. Additionally, we intro￾duce the Depth-Guided Boundary Loss (DGBL) to guide the model’s boundary learning on unlabeled data. Experimental results demonstrate that Dep… view at source ↗
Figure 2
Figure 2. Qualitative comparison on the KITTI Semantics dataset. The results [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 39 canonical work pages

  1. [29]

    Complementary random masking for rgb-thermal seman- tic segmentation,

    U. Shin et al., “Complementary random masking for rgb-thermal seman- tic segmentation,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 11 110–11 117

    work page 2024

  2. [1]

    Lightweight dual stream network with knowledge distillation for rgb-d scene parsing,

    Y . Zhang et al. , “Lightweight dual stream network with knowledge distillation for rgb-d scene parsing,” IEEE Signal Processing Letters , vol. 31, pp. 855–859, 2024

    work page 2024

  3. [2]

    SNE-RoadSeg: Incorporating Surface Normal Information into Semantic Segmentation for Accurate Freespace Detection,

    R. Fan et al., “SNE-RoadSeg: Incorporating Surface Normal Information into Semantic Segmentation for Accurate Freespace Detection,” in European Conference on Computer Vision (ECCV) . Springer, 2020, pp. 340–356

    work page 2020

  4. [3]

    S 3M-Net: Joint learning of semantic segmentation and stereo matching for autonomous driving,

    Z. Wu et al. , “S 3M-Net: Joint learning of semantic segmentation and stereo matching for autonomous driving,” IEEE Transactions on Intelligent Vehicles, vol. 9, no. 2, pp. 3940–3951, 2024

    work page 2024

  5. [4]

    SNE-RoadSegV2: Advancing heterogeneous feature fusion and fallibility awareness for freespace detection,

    Y . Feng et al. , “SNE-RoadSegV2: Advancing heterogeneous feature fusion and fallibility awareness for freespace detection,” IEEE Trans- actions on Instrumentation and Measurement , vol. 74, pp. 1–9, 2025

    work page 2025

  6. [5]

    Road damage detection based on unsupervised disparity map segmentation,

    R. Fan and M. Liu, “Road damage detection based on unsupervised disparity map segmentation,” IEEE Transactions on Intelligent Trans- portation Systems, vol. 21, no. 11, pp. 4906–4911, 2020

    work page 2020

  7. [6]

    Fully Convolutional Networks for Semantic Segmenta- tion,

    J. Long et al., “Fully Convolutional Networks for Semantic Segmenta- tion,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2015, pp. 3431–3440

    work page 2015

  8. [7]

    Pothole detection based on disparity transformation and road surface modeling,

    R. Fan et al. , “Pothole detection based on disparity transformation and road surface modeling,” IEEE Transactions on Image Processing , vol. 29, pp. 897–908, 2020

    work page 2020

Show all 46 references
  1. [8]

    RoadFormer+: Delivering RGB-X scene parsing through scale-aware information decoupling and advanced heteroge- neous feature fusion,

    J. Huang et al. , “RoadFormer+: Delivering RGB-X scene parsing through scale-aware information decoupling and advanced heteroge- neous feature fusion,” IEEE Transactions on Intelligent Vehicles , 2024, DOI: 10.1109/TIV .2024.3448251

    2024

  2. [9]

    MFFENet: Multiscale feature fusion and enhancement network for rgb–thermal urban road scene parsing,

    W. Zhou et al., “MFFENet: Multiscale feature fusion and enhancement network for rgb–thermal urban road scene parsing,” IEEE Transactions on Multimedia, vol. 24, pp. 2526–2538, 2021

    2021

  3. [10]

    MDNet: Mamba-effective diffusion-distillation network for rgb-thermal urban dense prediction,

    W. Zhou et al., “MDNet: Mamba-effective diffusion-distillation network for rgb-thermal urban dense prediction,” IEEE Transactions on Circuits and Systems for Video Technology , 2024

    2024

  4. [11]

    FuseNet: Incorporating Depth into Semantic Seg- mentation via Fusion-Based CNN Architecture,

    C. Hazirbas et al. , “FuseNet: Incorporating Depth into Semantic Seg- mentation via Fusion-Based CNN Architecture,” in 13th Asian Confer- ence on Computer Vision (ACCV) . Springer, 2017, pp. 213–228

    2017

  5. [12]

    RoadFormer: Duplex Transformer for RGB-Normal Se- mantic Road Scene Parsing,

    J. Li et al. , “RoadFormer: Duplex Transformer for RGB-Normal Se- mantic Road Scene Parsing,” IEEE Transactions on Intelligent Vehicles, 2024, DOI:10.1109/TIV .2024.3388726

    2024

  6. [13]

    Efficient multimodal semantic segmentation via dual- prompt learning,

    S. Dong et al. , “Efficient multimodal semantic segmentation via dual- prompt learning,” arXiv preprint arXiv:2312.00360 , 2023

    2023 arXiv

  7. [14]

    Depth-assisted semi-supervised rgb-d rail surface defect inspection,

    J. Wang et al., “Depth-assisted semi-supervised rgb-d rail surface defect inspection,” IEEE Transactions on Intelligent Transportation Systems , 2024

    2024

  8. [15]

    Unimatch v2: Pushing the limit of semi-supervised semantic segmentation,

    L. Yang et al. , “Unimatch v2: Pushing the limit of semi-supervised semantic segmentation,” arXiv preprint arXiv:2410.10777 , 2024

    2024 arXiv

  9. [16]

    Semi-supervised semantic segmentation needs strong, high-dimensional perturbations,

    G. French, T. Aila, S. Laine, M. Mackiewicz, and G. Finlayson, “Semi-supervised semantic segmentation needs strong, high-dimensional perturbations,” 2019

    2019

  10. [17]

    Improved regularization of convolutional neural networks with cutout,

    T. DeVries, “Improved regularization of convolutional neural networks with cutout,” arXiv preprint arXiv:1708.04552 , 2017

    2017 arXiv

  11. [18]

    Cutmix: Regularization strategy to train strong classifiers with localizable features,

    S. Yun et al., “Cutmix: Regularization strategy to train strong classifiers with localizable features,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 6023–6032

    2019

  12. [19]

    Revisiting weak-to-strong consistency in semi- supervised semantic segmentation,

    L. Yang et al. , “Revisiting weak-to-strong consistency in semi- supervised semantic segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 7236–7246

    2023

  13. [20]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence,

    K. Sohn et al. , “Fixmatch: Simplifying semi-supervised learning with consistency and confidence,” Advances in neural information processing systems, vol. 33, pp. 596–608, 2020

    2020

  14. [21]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab et al. , “Dinov2: Learning robust visual features without supervision,” Transactions on Machine Learning Research , 2023

    2023

  15. [22]

    Vision transformers for dense prediction,

    R. Ranftl et al. , “Vision transformers for dense prediction,” in Pro- ceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 12 179–12 188

    2021

  16. [23]

    Self-enhanced feature fusion for rgb-d semantic segmentation,

    P. Xiang et al. , “Self-enhanced feature fusion for rgb-d semantic segmentation,” IEEE Signal Processing Letters , 2024

    2024

  17. [24]

    DFormer: Rethinking RGBD representation learning for semantic segmentation,

    Y . Bowen et al. , “DFormer: Rethinking RGBD representation learning for semantic segmentation,” in International Conference on Learning Representations (ICLR), 2024

    2024

  18. [25]

    Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,

    A. Tarvainen and H. Valpola, “Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,” Advances in neural information processing systems , vol. 30, 2017

    2017

  19. [26]

    Frnet: Feature reconstruction network for rgb-d indoor scene parsing,

    W. Zhou et al., “Frnet: Feature reconstruction network for rgb-d indoor scene parsing,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 4, pp. 677–687, 2022

    2022

  20. [27]

    Feature contrast difference and enhanced network for RGB-D indoor scene classification in internet of things,

    W. Zhou et al., “Feature contrast difference and enhanced network for RGB-D indoor scene classification in internet of things,” IEEE Internet of Things Journal , 2025

    2025

  21. [28]

    Continuous pseudo-label rectified domain adaptive semantic segmentation with implicit neural representations,

    R. Gong et al. , “Continuous pseudo-label rectified domain adaptive semantic segmentation with implicit neural representations,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7225–7235

    2023

  22. [30]

    Self-supervised model adaptation for multimodal semantic segmentation,

    A. Valada et al. , “Self-supervised model adaptation for multimodal semantic segmentation,” International Journal of Computer Vision , vol. 128, no. 5, pp. 1239–1285, 2020

    2020

  23. [31]

    Active boundary loss for semantic segmentation,

    C. Wang et al. , “Active boundary loss for semantic segmentation,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 36, no. 2, 2022, pp. 2397–2405

    2022

  24. [32]

    Conditional boundary loss for semantic segmentation,

    D. Wu et al. , “Conditional boundary loss for semantic segmentation,” IEEE Transactions on Image Processing , 2023

    2023

  25. [33]

    Guided contrastive boundary learning for semantic segmentation,

    S. Qiu et al. , “Guided contrastive boundary learning for semantic segmentation,” Pattern Recognition, vol. 155, p. 110723, 2024

    2024

  26. [34]

    Effective whole-body pose estimation with two-stages distillation,

    Z. Yang et al. , “Effective whole-body pose estimation with two-stages distillation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 4210–4220

    2023

  27. [35]

    Augmented reality meets computer vision: Efficient data generation for urban driving scenes,

    A. Alhaija et al. , “Augmented reality meets computer vision: Efficient data generation for urban driving scenes,” International Journal of Computer Vision, vol. 126, pp. 961–972, 2018

    2018

  28. [36]

    Playing to Vision Foundation Model’s Strengths in Stereo Matching,

    C.-W. Liu et al. , “Playing to Vision Foundation Model’s Strengths in Stereo Matching,” IEEE Transactions on Intelligent Vehicles, 2024, DOI: 10.1109/TIV .2024.3467287

    2024

  29. [37]

    Decoupled Weight Decay Regularization,

    I. Loshchilov and F. Hutter, “Decoupled Weight Decay Regularization,” in International Conference on Learning Representations (ICLR) , 2019

    2019

  30. [38]

    DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs,

    L. Chen et al. , “DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 40, no. 4, pp. 834–848, 2017

    2017

  31. [39]

    Acnet: Attention based network to exploit complementary features for rgbd semantic segmentation,

    X. Hu et al., “Acnet: Attention based network to exploit complementary features for rgbd semantic segmentation,” in 2019 IEEE international conference on image processing (ICIP) . IEEE, 2019, pp. 1440–1444

    2019

  32. [40]

    CMX: Cross-modal fusion for RGB-X semantic segmentation with Transformers,

    J. Zhang et al. , “CMX: Cross-modal fusion for RGB-X semantic segmentation with Transformers,” IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 12, pp. 14 679–14 694, 2023

    2023

  33. [41]

    Sigma: Siamese mamba network for multi-modal semantic segmentation,

    Z. Wan et al. , “Sigma: Siamese mamba network for multi-modal semantic segmentation,” arXiv preprint arXiv:2404.04256 , 2024

    2024 arXiv

  34. [42]

    Geminifusion: Efficient pixel-wise multimodal fusion for vision transformer,

    D. Jia et al., “Geminifusion: Efficient pixel-wise multimodal fusion for vision transformer,” in Forty-first International Conference on Machine Learning, 2024

    2024

  35. [43]

    Sun RGB-D: A RGB-D scene understanding benchmark suite,

    S. Song et al., “Sun RGB-D: A RGB-D scene understanding benchmark suite,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 567–576

    2015

  36. [44]

    Improving semantic segmentation via video propagation and label relaxation,

    Y . Zhu et al., “Improving semantic segmentation via video propagation and label relaxation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2019, pp. 8856– 8865

    2019

  37. [45]

    Multi-target pan-class intrinsic relevance driven model for improving semantic segmentation in autonomous driving,

    Y . Cai et al. , “Multi-target pan-class intrinsic relevance driven model for improving semantic segmentation in autonomous driving,” IEEE Transactions on Image Processing , vol. 30, pp. 9069–9084, 2021

    2021

  38. [46]

    Warp-refine propagation: Semi-supervised auto- labeling via cycle-consistency,

    A. Ganeshan et al. , “Warp-refine propagation: Semi-supervised auto- labeling via cycle-consistency,” in Proceedings of the IEEE/CVF inter- national conference on computer vision , 2021, pp. 15 499–15 509

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.