URNet: A Unified Reparameterized Network for Efficient RGB-D Semantic Segmentation

Dongchen Zhu; Guangwei Gao; Guoan Xu; Ligeng Chen; Yang Xiao; Zhengxue Wang

REVIEW 4 major objections 4 minor 41 references

URNet: A Unified Reparameterized Network for Efficient RGB-D Semantic Segmentation

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

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

Pith's one-line read URNet claims a single reparameterized encoder with linear gated attention can match or beat dual-encoder transformer models on RGB-D semantic segmentation while using roughly half the compute.

desk verdict A promising efficiency-focused RGB-D architecture, but the unstated depth-source for ImageNet pretraining is a load-bearing omission that blocks verification. read the letter →

arxiv 2608.05671 v1 pith:4VLDLVT5 submitted 2026-08-06 cs.CV

Guoan Xu , Zhengxue Wang , Yang Xiao , Ligeng Chen , Guangwei Gao , Dongchen Zhu This is my paper · ORCID

classification cs.CV

keywords RGB-Dsemanticsegmentationreparameterizationcross-modalfusionlineargatedattentionpyramidmergingdecoderunifiedpretrainingefficientinferencesalientobjectdetection

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

URNet claims that RGB-D semantic segmentation does not need two separate encoders. The paper proposes a single-encoder network whose every block performs both multi-modal feature extraction and cross-modal fusion, using reparameterized convolutions for fast inference and a linear gated attention mechanism for RGB-depth interaction. On NYUDepth V2 and SUN-RGBD, URNet-L reports 58.8% and 53.4% mIoU at 73.0G FLOPs, outperforming DFormerV2-L (58.4% and 53.3% at 124.1G FLOPs). A large-scale RGB-D ImageNet-1k pretraining phase is the key enabler. If the claims hold, cheaper and faster RGB-D perception becomes available for robotics and mobile devices.

What carries the argument

The load-bearing mechanism is the Reparameterized Block (RepBlock): at training time it runs three parallel depthwise-convolution branches (a 3x3, a 1x1, and a decomposed 1x7+7x1) plus a skip connection, all of which fold into a single 3x3 depthwise convolution at inference, removing the extra branches' cost. Inside each block, the Linear Gated Attention (LGA) module produces a channel-wise gating signal from each modality via pooling and a 3x3 convolution, multiplies it onto the opposite modality, SE-refines both, and concatenates them into the block's fused output. Around the encoder, the Pyramid Merging Decoder (PMD) downsamples and adds adjacent-scale features then applies pyramid pooling with a residual connection. The whole framework is trained in two phases, RGB-D ImageNet-1k pretraining then segmentation fine-tuning, which the paper argues removes the modality mismatch of RGB-only pretrained encoders.

What would settle it

Retrain URNet-S on NYUDepth V2 with the same recipe but three pretraining depth sources: (a) the undisclosed original source, (b) a public monocular depth estimator on ImageNet, and (c) real sensor depth from NYUDepth V2 training images. If (b) and (c) reproduce the reported +1.2 mIoU gap over RGB-only pretraining within 0.3 mIoU, the claim survives; if the gap collapses for either, the pretraining mechanism is falsified. Simpler: check whether Table 4's numbers reproduce when the pretraining depth is explicitly documented.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a unified RGB-D encoder, pretrained jointly on paired RGB and depth images and fused at every block, can match or surpass the accuracy of dual-encoder transformer models while consuming substantially less computation. Concretely, URNet-L achieves 58.8% mIoU on NYUDepth V2 and 53.4% on SUN-RGBD with 73.0G FLOPs, compared with DFormerV2-L's 58.4% and 53.3% at 124.1G FLOPs. The authors attribute this to three co-designed components: reparameterized token mixers that fold multi-branch convolutions into a single inference-time convolution, a Linear Gated Attention (LGA) fusion that re-weights each modality by the other's compact descriptor, and a Pyramid Merging Decoder (PMD) that merges multi-scale features top-down with pooling and addition. The same encoder also transfers to RGB-D salient object detection, where it reports competitive or better results than large saliency models at lower FLOPs.

Load-bearing premise

The load-bearing premise is that the RGB-D ImageNet-1k pretraining data, whose depth maps are never sourced (ImageNet has no real depth), matches the sensor depth distribution used at fine-tuning; if that match fails, the central pretraining benefit could weaken or disappear.

Editorial extensions

If this is right

A single-encoder RGB-D architecture removes the duplicate feature-extraction branch, so inference cost grows like one backbone instead of two, with reported FPS up to 133 for URNet-T.
Reparameterization means the training-time multi-branch structure costs nothing at deployment, yielding a roughly 5x speedup over a self-attention token mixer at comparable accuracy (100.35 vs 18.23 FPS).
RGB-D pretraining adds 1.2 mIoU over RGB-only pretraining under identical fine-tuning, implying the unified pretraining is doing real work.
At the Large scale, URNet-L beats DFormerV2-L on both benchmarks with nearly half the FLOPs (73.0G vs 124.1G on NYUDepth V2).
The same pretrained encoder transfers to RGB-D salient object detection, outperforming HiDANet (130.6M, 71.5G) with 62.0M and 22.3G.

Reading between the lines

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

Because ImageNet has no real depth maps, the pretraining depth must come from an undisclosed estimator or synthetic source; a direct test is to swap that source (e.g., monocular depth estimates vs real sensor depth from NYU) and re-measure the +1.2% pretraining gain.
The LGA cross-gating and PMD decoder are modality-agnostic enough that the same architecture could be applied to other RGB-X tasks (thermal, event, polarization), where a single-encoder design would likely show similar savings.
Table 8's encoder-decoder swaps suggest the encoder, not the decoder, drives most of the gain; a follow-up could ablate URNet's encoder with an even lighter decoder or a different fusion head to isolate where the remaining headroom 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.

Desk Editor's Note

read the letter

URNet is a sensible efficiency-focused architecture paper. What's actually new is the specific combination of a reparameterized single-encoder backbone (RepBlock), a lightweight cross-gating fusion (LGA), and a pyramid merging decoder (PMD), with results on NYUDepth V2, SUN-RGBD, and RGB-D SOD. The ablations are well organized and generally support the design choices: the reparameterized token mixer beats self-attention in speed with comparable accuracy, LGA outperforms add and concatenation, and PMD improves over MLP and HAM. The large-scale comparison is the strongest part: URNet-L gets 58.8% mIoU on NYU and 53.4% on SUN-RGBD at 73.0G FLOPs, versus DFormerV2-L at 58.4/53.3 with 124.1G. That is a real efficiency claim.

The soft spots are serious, though. The paper never says where the depth maps for 'RGB-D ImageNet-1k' pretraining come from. ImageNet has no real depth, so the pretraining must use estimated or synthetic depth. This is not a footnote; Table 4 credits RGB-D pretraining with +1.2 mIoU, and Table 2 compares against DFormer under the 'same RGB-D setting.' Without the depth source, the central mechanism is unverifiable, and there is a risk the depth estimator leaks information from the evaluation domains if it was trained on indoor RGB-D data. The authors need to disclose the depth generation protocol before the pretraining claims can be taken at face value.

There are also internal inconsistencies. The HAM FLOPs are 18.40G in Table 7 but the text says the rise goes to 28.40G. Parameter counts for URNet-B are 44.1M in Table 1, 43.9M in Table 3, and 40.5M in Table 8; some of that may be decoder-specific, but it is not explained. There are no error bars or multiple seeds, which matters when the margins are under a point.

The novelty is incremental rather than conceptual: LGA resembles earlier cross-gating like SA-Gate, RepBlock extends RepViT with a decomposed 7x7 branch, and PMD is a simple pyramid summation. That is acceptable if the efficiency story holds. The paper deserves a serious referee, but it needs major revision: disclose the depth source, correct the tables, and ideally release code. As it stands, I would not rely on the exact numbers, but the architecture is worth engaging with.

Referee Report

4 major / 4 minor

Summary. The paper proposes URNet, a unified single-encoder RGB-D semantic segmentation network. The architecture combines a reparameterized block (RepBlock) with a linear gated attention (LGA) fusion module and a pyramid merging decoder (PMD). The model is first pretrained on paired RGB-D ImageNet-1k data and then fine-tuned on NYUDepth V2, SUN-RGBD, and several RGB-D salient object detection benchmarks. The authors report accuracy-efficiency advantages over dual-encoder transformer baselines, particularly DFormer and DFormerV2, and ablate the token mixer, fusion strategy, and decoder design.

Significance. If the reported results are correct, URNet would be a meaningful step toward efficient RGB-D segmentation: a single-encoder, reparameterized backbone with channel-gating fusion can match or exceed the accuracy of dual-encoder transformer models at substantially lower FLOPs and higher throughput. The paper also shows a plausible path to unified RGB-D pretraining and provides systematic ablations of the main components. The claims are, however, conditional on an underspecified pretraining data-generation protocol, and several internal numerical inconsistencies make it difficult to fully assess the efficiency and accuracy comparisons.

major comments (4)

[Section 3.1 and Section 4.1] The Phase I pretraining uses an 'RGB-D ImageNet-1k' dataset, but ImageNet does not contain real depth maps and the paper never states how the paired depth images were obtained. This is load-bearing because Table 4 attributes a +1.2% mIoU gain to RGB-D pretraining, Figure 1 credits the unified pretraining for cross-modal quality, and Table 2 compares URNet with DFormer variants under a claimed 'same RGB-D setting.' If the depth maps come from a monocular estimator, from synthetic renderings, or from a sensor-based collection, the transfer behavior and the risk of leakage from indoor RGB-D domains differ materially. The authors must disclose the depth source, the generation algorithm, preprocessing, and any overlap between the pretraining data and the evaluation datasets, and ideally provide a robustness test with an alternative depth source.
[Section 4.3, Table 7] The text states that the HAM decoder raises FLOPs from 13.98G to 28.40G, but Table 7 reports 18.40G for HAM. This discrepancy directly affects the claimed efficiency advantage of PMD over HAM. The correct value must be reported consistently, and the ablation conclusion should be restated with the corrected numbers.
[Section 4.2, Table 3] Several headline comparisons rest on very small margins: URNet-S versus DFormerV2-S is 56.1% vs 56.0% on NYUDepth V2 and 51.8% vs 51.5% on SUN-RGBD, and URNet-L versus DFormerV2-L is 58.8% vs 58.4% and 53.4% vs 53.3%. No error bars, confidence intervals, or multiple-seed experiments are reported. Given that the paper claims state-of-the-art performance, the authors should report the number of runs and standard deviations for at least these key comparisons, or soften the claims accordingly.
[Tables 1, 3, and 8] Parameter counts are inconsistent across tables for the same model: URNet-T is listed as 9.2M in Table 1 but 9.3M in Table 3; URNet-B is 44.1M versus 43.9M; URNet-L is 62.2M versus 61.8M; and Table 8 reports URNet-B with PMD as 43.9M. The authors should specify whether decoder parameters are included in each count and reconcile the numbers, since the efficiency comparisons are central to the paper.

minor comments (4)

[Section 4.4] The paper refers to 'Supplements' for more ablation studies, but the submitted version contains no supplementary material; the file should be included or the mention removed.
[Table 3] The label 'URNet (ours) 26 UR' in the method rows is unclear; it appears to be an artifact and should be replaced with a clear 'Ours' marker.
[Table 9] DFormerV2-S and DFormerV2-L list '-' for parameter count and FLOPs, which makes the efficiency comparison incomplete; please provide these values or state why they are unavailable.
[Section 3.1] Equation (1) writes the reparameterized token mixer with four additive branches, but the inference-time equivalence to a single convolution is not explicitly shown; a short derivation or reference to RepViT would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: URNet's results are empirical evaluations against external benchmarks, and the unspecified RGB-D ImageNet depth source is a reproducibility gap, not a self-referential derivation.

full rationale

I walked the claimed derivation chain. URNet's contributions are architectural (RepBlock, LGA, PMD) and are evaluated by fine-tuning on external benchmarks (NYUDepth V2, SUN-RGBD, RGB-D SOD). No quantity is fitted to a benchmark and then reported as a prediction; no equation is defined in terms of the quantity it is said to derive. The central pretraining claim (Section 3.1 and Section 4.1) is conditional on an unstated depth-generation source, since ImageNet-1k contains no real depth maps; this is a reproducibility and validity gap, not circularity, because the RGB-versus-RGB-D pretraining comparison in Table 4 is an empirical ablation whose outcome is not encoded in the definition of either training scheme. The self-citations [31, 32, 33] appear only as related-work context for backbones and dual-encoder designs; they do not justify the novel fusion or decoder, and no uniqueness theorem or external authority is invoked to forbid alternatives. The comparisons to DFormer and DFormerV2 rely on independently published external work, and the reported numbers stand or fall on experimental reproducibility rather than on any self-referential chain. Therefore no circular step meets the quoted-reduction standard.

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

The central claims rest on hand-chosen architectural hyperparameters, the existence and validity of paired ImageNet depth data, and the comparability of prior published numbers. No new physical or conceptual entities are introduced; LGA and PMD are architectural modules evaluated empirically.

free parameters (2)

Encoder stage channel widths and block counts (Table 1) = URNet-T/S/B/L: e.g., stage1 channels 48/56/80/80; blocks 3/5/7/7
Hand-chosen architectural hyperparameters following DFormer and RepViT conventions; no sensitivity analysis is reported, and the specific values are not derived from any principle.
Decoder hidden dimension = 256 for all scales
Fixed by hand for all model sizes; the paper does not test alternative decoder widths or report how this choice affects the accuracy-efficiency balance.

assumptions (4)

domain assumption Paired RGB-D ImageNet-1k data exists and is usable for pretraining
Sections 3.1 and 4.1 assume RGB-D ImageNet-1k pretraining; ImageNet has no real depth, and the paper does not state how depth is obtained, leaving a load-bearing protocol unspecified.
standard math Multi-branch depthwise convolutions can be folded into a single convolution at inference
Equation (1) sums depthwise convolutions and a skip connection; the reparameterization claim relies on linearity, a standard but unproven assumption in this paper.
domain assumption Channel-level statistics from each modality are sufficient for effective cross-modal gating
LGA (Equations 2-5) pools each feature to a global descriptor and gates the other modality; this assumes channel statistics carry the complementary information needed for fusion.
domain assumption Reported results from prior methods were produced under comparable training settings
Table 3 compares with numbers taken from original papers; there is no reimplementation under identical fine-tuning schedules, resolutions, or augmentations, so small differences could affect the comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of URNet: A Unified Reparameterized Network for Efficient RGB-D Semantic Segmentation." pith.science (2026). https://pith.science/paper/4VLDLVT5

@misc{pith2026260805671,
  author       = {Pith},
  title        = {Pith review of: URNet: A Unified Reparameterized Network for Efficient RGB-D Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4VLDLVT5}},
  note         = {Machine review of arXiv:2608.05671}
}

read the original abstract

Previous RGB-D semantic segmentation methods commonly employ dual encoders to separately process RGB and depth inputs, followed by dedicated modules for cross-modal feature fusion. However, such designs often inadequately capture depth representations and consequently limit effective cross-modal interaction, while the additional encoder branch introduces redundant computation that hinders lightweight execution. To tackle these challenges, we propose URNet, a Unified Reparameterized RGB-D Network that performs simultaneous multi-modal feature extraction and cross-modal fusion within a single encoder. Specifically, we adopt a reparameterization strategy to compact the network architecture and facilitate fast inference. Within each Reparameterized Block (RepBlock), a Linear Gated Attention (LGA) module is introduced to fully exploit complementary RGB and depth cues across different feature scales. Furthermore, considering that decoder design has been relatively underexplored in existing RGB-D segmentation models, we develop a concise yet effective universal decoder, termed the Pyramid Merging Decoder (PMD). Extensive experiments on multiple RGB-D segmentation benchmarks demonstrate that URNet achieves state-of-the-art performance while maintaining high efficiency. Code will be available at https://github.com/Wild-Stephen/URNet.

Figures

Figures reproduced from arXiv: 2608.05671 by the authors.

**Figure 3.** Detailed structure of the RepBlock with Linear Gated Attention (LGA). [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Structural details of the Pyramid Merging Decoder [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (4 more)

**Figure 5.** Figure 5: Qualitative comparisons on NYU-Depth V2. From left to right are RGB images, depth images, and segmentation [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Qualitative comparisons on SUN RGB-D. From left to right are RGB images, depth images, and segmentation results [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Visualizations under different fusion strategies. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 9.** Figure 9: Qualitative Results of Salient Object Detection. From left to right are RGB images, depth images, ground truth and [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 31 canonical work pages

[1]

Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. 2021. Xcit: Cross-covariance image transformers.Advances in Neural Information Processing Systems34 (2021), 20014–20027

work page 2021
[2]

Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. 2022. Multi- mae: Multi-modal multi-task masked autoencoders. InProceedings of the European Conference on Computer Vision. Springer, 348–367

work page 2022
[3]

Lizhi Bai, Jun Yang, Chunqi Tian, Yaoru Sun, Maoyu Mao, Yanjun Xu, and Weirong Xu. 2025. DCANet: Differential convolution attention network for RGB-D se- mantic segmentation.Pattern Recognition162 (2025), 111379

work page 2025
[4]

Jinming Cao, Hanchao Leng, Dani Lischinski, Daniel Cohen-Or, Changhe Tu, and Yangyan Li. 2021. Shapeconv: Shape-aware convolutional layer for indoor rgb-d semantic segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 7088–7097

work page 2021
[5]

Lin-Zhuo Chen, Zheng Lin, Ziqin Wang, Yong-Liang Yang, and Ming-Ming Cheng

work page
[6]

Qian Chen, Zhenxi Zhang, Yanye Lu, Keren Fu, and Qijun Zhao. 2022. 3-D convolutional neural networks for RGB-D salient object detection and beyond. IEEE Transactions on Neural Networks and Learning Systems35, 3 (2022), 4309– 4323

work page 2022
[7]

Xiaokang Chen, Kwan-Yee Lin, Jingbo Wang, Wayne Wu, Chen Qian, Hongsheng Li, and Gang Zeng. 2020. Bi-directional cross-modality feature propagation with separation-and-aggregation gate for RGB-D semantic segmentation. In Proceedings of the European Conference on Computer Vision. Springer, 561–577

work page 2020
[8]

Shaohua Dong, Yunhe Feng, Qing Yang, Yan Huang, Dongfang Liu, and Heng Fan. 2024. Efficient multimodal semantic segmentation via dual-prompt learning. InProceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 14196–14203

work page 2024

Show all 41 references

[9]

Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

2020 arXiv
[10]

Siqi Du, Weixi Wang, Renzhong Guo, Ruisheng Wang, and Shengjun Tang. 2024. Asymformer: Asymmetrical cross-modal representation learning for mobile plat- form real-time rgb-d semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

2024
[11]

Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens Van Der Maaten, Armand Joulin, and Ishan Misra. 2022. Omnivore: A single model for many visual modali- ties. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 16102–16112

2022
[12]

Yan Gong, Jianli Lu, Yongsheng Gao, Jie Zhao, Xiaojuan Zhang, and Susanto Rahardja. 2025. DiffPixelFormer: Differential Pixel-Aware Transformer for RGB-D Indoor Scene Segmentation.arXiv preprint arXiv:2511.13047(2025)

2025
[13]

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 770–778

2016
[14]

Ding Jia, Jianyuan Guo, Kai Han, Han Wu, Chao Zhang, Chang Xu, and Xinghao Chen. 2024. GeminiFusion: Efficient Pixel-wise Multimodal Fusion for Vision Transformer. arXiv:2406.01210 [cs.CV]

2024 arXiv
[15]

Di Jia, Chen Zhao, Huilun Song, Huaxiu Zhang, and Wei Li. 2025. ECMRN: Efficient Cross-Modal Reparameterization Network for RGB-D Tasks via Prompt Tuning.Knowledge-Based Systems(2025), 114321

2025
[16]

Minhyeok Lee, Chaewon Park, Suhwan Cho, and Sangyoun Lee. 2022. Spsn: Superpixel prototype sampling network for rgb-d salient object detection. In Proceedings of the European Conference on Computer Vision. Springer, 630–647

2022
[17]

Bingyu Li, Da Zhang, Zhiyuan Zhao, Junyu Gao, and Xuelong Li. 2025. Stitch- fusion: Weaving any visual modalities to enhance multimodal semantic seg- mentation. InProceedings of the ACM International Conference on Multimedia. 1308–1317

2025
[18]

Nian Liu, Ni Zhang, Kaiyuan Wan, Ling Shao, and Junwei Han. 2021. Visual saliency transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision. 4722–4732

2021
[19]

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF International Conference on Computer Vision. 10012–10022

2021
[20]

Sachin Mehta and Mohammad Rastegari. 2021. Mobilevit: light-weight, general- purpose, and mobile-friendly vision transformer.arXiv preprint arXiv:2110.02178 (2021)

2021 arXiv
[21]

Daniel Seichter, Söhnke Benedikt Fischedick, Mona Köhler, and Horst-Michael Groß. 2022. Efficient multi-task rgb-d scene analysis for indoor environments. InProceedings of the International Joint Conference on Neural Networks (IJCNN). IEEE, 1–10

2022
[22]

Daniel Seichter, Mona Köhler, Benjamin Lewandowski, Tim Wengefeld, and Horst-Michael Gross. 2021. Efficient rgb-d semantic segmentation for indoor scene analysis. InProceedings of the IEEE International Conference on Robotics and Automation (ICRA). IEEE, 13525–13531

2021
[23]

Zifu Wan, Pingping Zhang, Yuhao Wang, Silong Yong, Simon Stepputtis, Katia Sycara, and Yaqi Xie. 2025. Sigma: Siamese mamba network for multi-modal semantic segmentation. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). IEEE, 1734–1744

2025
[24]

Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. 2024. Repvit: Revisiting mobile cnn from vit perspective. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. 15909–15920

2024
[25]

Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. 2021. Pyramid vision transformer: A versatile back- bone for dense prediction without convolutions. InProceedings of the IEEE/CVF International Conference on Computer Vis...

2021
[26]

Yikai Wang, Xinghao Chen, Lele Cao, Wenbing Huang, Fuchun Sun, and Yunhe Wang. 2022. Multimodal token fusion for vision transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 12186–12195

2022
[27]

Yikai Wang, Wenbing Huang, Fuchun Sun, Tingyang Xu, Yu Rong, and Junzhou Huang. 2020. Deep multimodal fusion by channel exchanging.Advances in Neural Information Processing Systems33 (2020), 4835–4845

2020
[28]

Zongwei Wu, Guillaume Allibert, Fabrice Meriaudeau, Chao Ma, and Cédric Demonceaux. 2023. Hidanet: Rgb-d salient object detection via hierarchical depth awareness.IEEE Transactions on Image Processing32 (2023), 2160–2173

2023
[29]

Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. 2021. SegFormer: Simple and efficient design for semantic segmentation with transformers.Advances in Neural Information Processing Systems34 (2021), 12077–12090

2021
[30]

Cunlu Xu, Gang Ma, Feng Gao, Bin Wang, and Jun Liu. 2025. ADBNet: Asym- metric dual-branch network for indoor real-time RGB-D semantic segmentation. Knowledge-Based Systems326 (2025), 113885

2025
[31]

Guoan Xu, Jiaming Chen, Wenfeng Huang, Wenjing Jia, Guangwei Gao, and Guo- Jun Qi. 2026. SCASeg: Strip Cross-Attention for Efficient Semantic Segmentation. IEEE Transactions on Image Processing35 (2026), 4847–4860

2026
[32]

Guoan Xu, Wenfeng Huang, Wenjing Jia, Jiamao Li, Guangwei Gao, and Guo-Jun Qi. 2025. S2AFormer: Strip Self-Attention for Efficient Vision Transformer.IEEE Transactions on Image Processing34 (2025), 8243–8256

2025
[33]

Guoan Xu, Yang Xiao, Guangwei Gao, Dongchen Zhu, Guo-Jun Qi, and Wen- jing Jia. 2026. RSGMamba: Reliability-Aware Self-Gated State Space Model for Multimodal Semantic Segmentation.arXiv preprint arXiv:2604.12319(2026)

2026 arXiv
[34]

Bowen Yin, Xuying Zhang, Zhong-Yu Li, Li Liu, Ming-Ming Cheng, and Qibin Hou. 2024. DFormer: Rethinking RGBD Representation Learning for Semantic Segmentation. InICLR

2024
[35]

Bo-Wen Yin, Jiao-Long Cao, Ming-Ming Cheng, and Qibin Hou. 2025. DFormerv2: Geometry Self-Attention for RGBD Semantic Segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Conference. 19345–19355

2025
[36]

Jiaming Zhang, Huayao Liu, Kailun Yang, Xinxin Hu, Ruiping Liu, and Rainer Stiefelhagen. 2023. CMX: Cross-modal fusion for RGB-X semantic segmentation with transformers.IEEE Transactions on intelligent transportation systems24, 12 (2023), 14679–14694

2023
[37]

Jiaming Zhang, Ruiping Liu, Hao Shi, Kailun Yang, Simon Reiß, Kunyu Peng, Haodong Fu, Kaiwei Wang, and Rainer Stiefelhagen. 2023. Delivering arbitrary- modal semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1136–1147

2023
[38]

Chuanyang Zheng. 2025. Iformer: Integrating ConvNet and transformer for mobile application.arXiv preprint arXiv:2501.15369(2025)

2025 arXiv
[39]

Tao Zhou, Huazhu Fu, Geng Chen, Yi Zhou, Deng-Ping Fan, and Ling Shao. 2021. Specificity-preserving RGB-D saliency detection. InProceedings of the IEEE/CVF International Conference on Computer Vision. 4681–4691

2021
[40]

Wujie Zhou, Enquan Yang, Jingsheng Lei, Jian Wan, and Lu Yu. 2022. PGDENet: Progressive guided fusion and depth enhancement network for RGB-D indoor scene parsing.IEEE Transactions on Multimedia25 (2022), 3483–3494

2022
[2021]

Spatial information guided convolution for real-time RGBD semantic segmentation.IEEE Transactions on Image Processing30 (2021), 2313–2324

2021

Pith tools

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

[1] [1]

Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. 2021. Xcit: Cross-covariance image transformers.Advances in Neural Information Processing Systems34 (2021), 20014–20027

work page 2021

[2] [2]

Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. 2022. Multi- mae: Multi-modal multi-task masked autoencoders. InProceedings of the European Conference on Computer Vision. Springer, 348–367

work page 2022

[3] [3]

Lizhi Bai, Jun Yang, Chunqi Tian, Yaoru Sun, Maoyu Mao, Yanjun Xu, and Weirong Xu. 2025. DCANet: Differential convolution attention network for RGB-D se- mantic segmentation.Pattern Recognition162 (2025), 111379

work page 2025

[4] [4]

Jinming Cao, Hanchao Leng, Dani Lischinski, Daniel Cohen-Or, Changhe Tu, and Yangyan Li. 2021. Shapeconv: Shape-aware convolutional layer for indoor rgb-d semantic segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 7088–7097

work page 2021

[5] [5]

Lin-Zhuo Chen, Zheng Lin, Ziqin Wang, Yong-Liang Yang, and Ming-Ming Cheng

work page

[6] [6]

Qian Chen, Zhenxi Zhang, Yanye Lu, Keren Fu, and Qijun Zhao. 2022. 3-D convolutional neural networks for RGB-D salient object detection and beyond. IEEE Transactions on Neural Networks and Learning Systems35, 3 (2022), 4309– 4323

work page 2022

[7] [7]

Xiaokang Chen, Kwan-Yee Lin, Jingbo Wang, Wayne Wu, Chen Qian, Hongsheng Li, and Gang Zeng. 2020. Bi-directional cross-modality feature propagation with separation-and-aggregation gate for RGB-D semantic segmentation. In Proceedings of the European Conference on Computer Vision. Springer, 561–577

work page 2020

[8] [8]

Shaohua Dong, Yunhe Feng, Qing Yang, Yan Huang, Dongfang Liu, and Heng Fan. 2024. Efficient multimodal semantic segmentation via dual-prompt learning. InProceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 14196–14203

work page 2024