HDBFormer: Efficient RGB-D Semantic Segmentation with A Heterogeneous Dual-Branch Framework

Binghua Su; Shuobin Wei; Zhengan Lu; Zhuang Zhou; Zizhao Yuan

REVIEW 5 major objections 7 minor 46 references

HDBFormer: Efficient RGB-D Semantic Segmentation with A Heterogeneous Dual-Branch Framework

T0 review · 5 major / 7 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read HDBFormer claims depth images carry simpler information than RGB, so a lightweight CNN depth encoder with an asymmetric fusion module can beat heavier transformer encoders, reaching 59.3 mIoU on NYUDepthv2 and 53.9 on SUN-RGBD.

desk verdict Useful incremental RGB-D segmentation architecture with competitive numbers, but the central 'depth is simpler' story is not isolated and the SOTA claim is a tie on NYU. read the letter →

arxiv 2504.13579 v1 pith:ADQI25LH submitted 2025-04-18 cs.CV

Shuobin Wei , Zhuang Zhou , Zhengan Lu , Zizhao Yuan , Binghua Su This is my paper · ORCID

classification cs.CV

keywords RGB-Dsemanticsegmentationheterogeneousdual-branchframeworklightweightdepthencodermodalityasymmetrycross-modalfusiondepthwiseseparableconvolutionindoorsceneparsingNYUDepthv2

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 tries to establish that in indoor RGB-D semantic segmentation, depth images should not be processed with the same heavy machinery as RGB. It claims depth features are intrinsically less complex, so a tiny hierarchical CNN encoder (LDFormer, 0.7M parameters) suffices, and the fusion module should treat depth as a supporting modality that bypasses global attention. On NYUDepthv2 the complete model reaches 59.3 mIoU and on SUN-RGBD 53.9 mIoU, above the published models compared in the paper. If the claim is right, practitioners can cut most of the cost of the depth branch without losing accuracy, and the design principle extends to other low-complexity auxiliary modalities.

What carries the argument

The load-bearing mechanism is the asymmetry itself, realized by two named components. LDFormer is a lightweight hierarchical encoder made of 3x3 depthwise convolutions, 1x1 pointwise convolutions, and max-pool downsampling, producing four feature scales that match the RGB branch. MIIM is the fusion module: it splits each feature pair into primary and minor features, sends the pair through Global Fusion Attention, a Transformer whose query is pooled to 6x6 to cut cost, and Local Fusion Attention, a 7x7 large-kernel-convolution gate over the minor features, then iterates the interaction twice. Depth-as-primary features go only through Local Fusion Attention. That graded, asymmetric routing is what carries the efficiency and the accuracy claim.

What would settle it

A decisive check is to run the identical HDBFormer with the depth branch replaced by a standard 0.7M-parameter CNN of the same depth but without depthwise separable convolutions, holding everything else fixed; if it matches 59.3 mIoU on NYUDepthv2, the specific lightweight design is not what carries the result. Likewise, replacing the depth map with spatially shuffled patches would show whether the encoder is actually reading depth structure.

Watch

Extended reading notes

Core claim

The central discovery is an asymmetric treatment: RGB goes through a two-stream encoder (a deep Swin Transformer plus a shallow convolutional branch) while depth goes through LDFormer, a depthwise-separable-convolution hierarchical encoder. In the Modality Information Interaction Module, RGB-depth feature pairs pass through global fusion attention and local fusion attention, but depth-only features skip global attention; removing the depth-side global branch gives the best configuration, 59.3 mIoU versus 58.3 with all four MIIM components. LDFormer (0.7M parameters, 2.9G FLOPs) also beats Swin-T (27.5M, 28.6G) and ResNet-50 (11.7M, 25.3G) as the depth encoder inside the same system. The paper takes this as evidence that simpler encoders are not merely cheaper for depth, but more accurate.

Load-bearing premise

The load-bearing premise is that depth images carry less complex information than RGB images, so a tiny convolutional encoder and a depth path that skips global attention lose nothing; if depth features are actually as complex as RGB features, the reported gains could come from the fusion module or from architecture choice rather than from the asymmetry itself.

Editorial extensions

If this is right

On NYUDepthv2, swapping the depth encoder from Swin-T to LDFormer raises mIoU from 58.4 to 59.3 while cutting parameters from 27.5M to 0.7M and FLOPs from 28.6G to 2.9G.
Dropping the depth-side global fusion branch yields the best configuration, 59.3 mIoU versus 58.3 with all four MIIM components, so global attention on depth is not needed.
With the full model, reported mIoU is 59.3 on NYUDepthv2 and 53.9 on SUN-RGBD, above the 15 published models the paper tabulates.
A hierarchical CNN with depthwise separable convolutions is enough to build multi-scale depth features; transformer self-attention is unnecessary for the depth branch in this design.

Reading between the lines

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

If the asymmetry principle transfers, other auxiliary cues with simple spatial structure, such as surface normals, edge maps, or thermal images, could use the same cheap-encoder-plus-local-only-fusion recipe, but the paper only reports RGB-D and lists RGB-T as future work.
The near-tie between LDFormer (59.3) and ResNet-50 (59.0) as depth encoders leaves open whether the gain comes from the simplicity premise or simply from having fewer parameters to overfit on the 795-image NYUDepthv2 training set.
A sharper test of the premise would be to feed the same architecture a depth map with patches shuffled; if the lightweight encoder still holds its advantage, the benefit is parameter reduction rather than reading spatial layout.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Short version: this is a useful incremental RGB-D segmentation paper with competitive numbers, but the main story—that depth features are simpler and therefore need a lightweight encoder—is not actually isolated by the experiments. The SOTA claim on NYU is also a tie, not a win.

What's new and good. The architecture is reasonable: a two-stream RGB branch (Swin-B plus a small conv net) and a 0.7M-parameter depthwise-separable depth encoder, fused through a module that routes depth features only through local attention and keeps global attention for RGB. That asymmetry is the interesting idea, and it's cheap: 2.9G FLOPs for the depth encoder. The numbers on SUN-RGBD (53.9 mIoU) are the best in their own table, and on NYU they match DPLNet at 59.3. They also release code, and the ablations (Tables II and III in the main text) show the design choices affect performance.

Where it's soft. The biggest gap is the missing control for the central asymmetry claim. Table II varies the depth encoder while keeping the RGB branch fixed, but never swaps which modality gets the cheap encoder. So the 0.9-point difference between LDFormer and Swin-T as depth encoder could come from having one strong and one weak branch, from the fusion module, or from test-set-driven selection—not specifically from matching complexity to modality. That's a legitimate concern, not a nitpick.

Also: no error bars or multiple seeds anywhere; a 0.9-point gain on a single run is within the noise you typically see in this benchmark. The "state-of-the-art" claim in the abstract is overstated because it's a tie on NYU. And the supplementary has a direct contradiction: the text says MIIM outperforms other interaction modules in parameters and FLOPs, but Table I in the supplement shows iAFF has 2.8M params and 2.4G FLOPs versus MIIM's 10.2M and 9.7G. That claim is just wrong as written.

Who it's for. If you work on RGB-D segmentation and want a lightweight depth encoder that works with a transformer-based RGB branch, this is a usable reference point. It's not a conceptual breakthrough, and the motivating premise is under-supported, but the engineering is coherent and the numbers are competitive.

My take: it deserves a serious referee. The fixes are straightforward—add the swapped-complexity ablation, report multiple runs, correct the supplementary text—and the paper would be stronger for them. I'd send it to review rather than desk-reject.

Referee Report

5 major / 7 minor

Summary. The paper proposes HDBFormer, a heterogeneous dual-branch framework for RGB-D semantic segmentation. RGB features are extracted by a Swin Transformer detail encoder combined with a simple convolutional base encoder, while depth features are extracted by LDFormer, a lightweight hierarchical depthwise separable CNN. A Modality Information Interaction Module (MIIM) fuses the two modalities using a pooled-query global fusion attention (GFA) and a local fusion attention (LFA) based on large-kernel convolutions, with an iterative enhancement strategy (N=2). Experiments on NYUDepthv2 and SUN-RGBD report 59.3% and 53.9% mIoU respectively, matching or exceeding previous state-of-the-art results.

Significance. The paper's efficiency-oriented design for the depth branch (LDFormer: 0.7M parameters, 2.9G FLOPs) is a practical strength, and the release of code aids reproducibility. If the central asymmetry claim—that depth features are less complex than RGB features—were successfully isolated, the design principle of matching encoder complexity to modality would be a useful contribution to multimodal architecture design. However, the current experiments do not yet isolate this claim from confounds, and several methodological details are missing. The benchmark numbers themselves are competitive and provide a solid empirical baseline for the proposed architecture.

major comments (5)

[Section II-B, Tables II and III] The paper's central premise—that depth features are less complex than RGB features—is supported only by ablations that keep the heavy Swin-B encoder on RGB and LDFormer on depth. Table II shows LDFormer reaching 59.3 mIoU versus 58.4 for Swin-T as the depth encoder, and Table III shows the best configuration removing GFA2. Neither experiment varies which modality receives the lightweight encoder, so the observed gains could stem from the specific inductive bias of the depthwise separable CNN, from the fusion module, or from optimization dynamics rather than from matching model complexity to modality. A swapped-complexity ablation (e.g., LDFormer on RGB and Swin-T on depth, plus symmetric same-encoder baselines) is needed to attribute the improvement to the modality-complexity asymmetry.
[Section III-B and III-D] All ablations and architecture-selection decisions (choosing N=2, removing GFA2, selecting LDFormer over other depth encoders) are made using the test split of NYUDepthv2, and only single-run numbers are reported without error bars. Because the test set is used for component selection, the reported gains may partially reflect fitting to the benchmark, and the 0.9-point gap over Swin-T in Table II could lie within run-to-run variance. The paper should report mean and standard deviation over multiple seeds and should perform ablation-based model selection on a validation split, reserving the test set for a single final evaluation.
[Section II-B, Eq. (8)] The iterative enhancement strategy with N=2 is described in prose but not formalized. Equations (4)–(7) define only a single pass through MIIM; there is no equation stating whether the second iteration takes the outputs F'RGB_i and F'Depth_i from the first pass as inputs, how the loop is initialized, or how the decoder consumes the depth-side features. The decoder in Eq. (8) upsamples only F'RGB_1..4 and F_Base_4, so the role of F'Depth_i in the final prediction is unclear. The paper should specify the recurrence and the complete computational graph for the iterative procedure.
[Supplementary Table I and text] The supplementary text states that MIIM 'significantly outperforms existing modality information interaction module in terms of mIoU, parameters, and FLOPs,' yet the immediately preceding sentence acknowledges that iAFF has lower parameters (2.8M) and FLOPs (2.4G) than MIIM (10.2M, 9.7G). This is an internal contradiction: MIIM does not outperform iAFF on parameters or FLOPs. The claim should be corrected to state superiority in mIoU with competitive efficiency, not Pareto dominance.
[Section III-C, Table I, Abstract] The abstract claims state-of-the-art performance on NYUDepthv2, but Table I shows DPLNet also reporting 59.3 mIoU on that dataset, tying with HDBFormer. The SOTA claim should be qualified: HDBFormer matches DPLNet on NYUDepthv2 mIoU and leads on SUN-RGBD mIoU (53.9 versus 52.8).

minor comments (7)

[Eq. (2)] Equation (2) uses F_Depth_i in the RGB base/detail fusion formula, but the surrounding text describes fusing F_Detail_i (from Swin) with F_Base_i (from the simple encoder); this appears to be a typo and should read F_Detail_i.
[Eq. (7)] The two equations defining F'RGB_i and F'Depth_i are identical; if these outputs are intentionally the same, the paper should say so explicitly, otherwise the formula for F'Depth_i should be corrected.
[Eq. (6), Section II-B] The 'large kernel convolution' in LFA uses a 7×7 kernel, which is not typically considered 'large' (e.g., RepLKNet uses 31×31). Consider justifying the term or renaming the operation.
[Table III, main text] The checkmark notation in Table III is ambiguous; explicitly labeling the rows with the set of included components (e.g., 'LFA1+GFA1+LFA2') would improve readability.
[Section III-B] The implementation details omit the number of training epochs and the input resolution; these are standard hyperparameters and should be reported for reproducibility.
[Supplementary Table III] The 'w/o Pooling' row reports 36.127GB maximum memory, which exceeds the 24GB capacity of an RTX4090; the text should clarify that this configuration is infeasible on the used hardware and that the mIoU is therefore unavailable.
[Throughout] The paper does not report total parameters and FLOPs for the full HDBFormer model including the Swin-B and base encoders; such totals are necessary to substantiate the 'efficient' descriptor in the title and abstract.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's performance claims are validated against external benchmarks and its component choices are supported by ablations, not by equation-level reduction to their own inputs.

full rationale

HDBFormer is an empirical architecture paper. Its central claims are state-of-the-art mIoU on NYUDepthv2 and SUN-RGBD, measured against held-out test splits of external datasets, and the efficiency of its LDFormer depth encoder, measured by parameters, FLOPs, and mIoU in the same framework. There is no derivation chain in which an output quantity is defined in terms of the claimed prediction, no fitted parameter is relabeled as a prediction, and no load-bearing self-citation or imported uniqueness theorem appears. The motivating premise that depth features are less complex than RGB features is a design assumption, not a derived result; the ablations in Tables II and III provide empirical support for the chosen asymmetry, though they do not fully isolate the modality-complexity hypothesis from other confounds. That limitation is a correctness or experimental-design concern, not circularity under the stated rules. The absence of a separate validation set and the lack of error bars are also robustness concerns, but they do not make the reported benchmark results equivalent by construction to the paper's inputs. No equation in the paper reduces to another equation by definition, and no cited prior work by the same authors is used as the sole justification for a central premise. Accordingly, no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new mathematical objects; its 'free parameters' are a few architectural hyperparameters. The key assumption is the modality-complexity claim, which drives the entire design.

free parameters (3)

N (iterative enhancement steps) = 2
Chosen by hand; controls how many times MIIM repeats fusion and affects final mIoU.
GFA query pooling size = 6x6
Chosen by hand; the query is pooled to 6x6 to reduce attention cost; value affects representation granularity.
LFA large kernel size = 7x7
Chosen by hand; used in the linear-then-convolution path for local feature capture.

assumptions (3)

domain assumption Depth image features are less complex than RGB features.
This is the paper's core premise (Section I) that justifies the lightweight depth encoder and asymmetric fusion; it is not measured or proven in the paper.
domain assumption Swin Transformer, depthwise separable convolutions, and large-kernel convolutions are effective and reliable building blocks.
The model is assembled from these prior modules without re-derivation; performance depends on their known properties.
domain assumption The standard train/test splits and evaluation protocol for NYUDepthv2 and SUN-RGBD yield meaningful comparisons.
The paper uses standard splits and metrics but does not report variance or statistical significance.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HDBFormer: Efficient RGB-D Semantic Segmentation with A Heterogeneous Dual-Branch Framework." pith.science (2026). https://pith.science/paper/ADQI25LH

@misc{pith2026250413579,
  author       = {Pith},
  title        = {Pith review of: HDBFormer: Efficient RGB-D Semantic Segmentation with A Heterogeneous Dual-Branch Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ADQI25LH}},
  note         = {Machine review of arXiv:2504.13579}
}

read the original abstract

In RGB-D semantic segmentation for indoor scenes, a key challenge is effectively integrating the rich color information from RGB images with the spatial distance information from depth images. However, most existing methods overlook the inherent differences in how RGB and depth images express information. Properly distinguishing the processing of RGB and depth images is essential to fully exploiting their unique and significant characteristics. To address this, we propose a novel heterogeneous dual-branch framework called HDBFormer, specifically designed to handle these modality differences. For RGB images, which contain rich detail, we employ both a basic and detail encoder to extract local and global features. For the simpler depth images, we propose LDFormer, a lightweight hierarchical encoder that efficiently extracts depth features with fewer parameters. Additionally, we introduce the Modality Information Interaction Module (MIIM), which combines transformers with large kernel convolutions to interact global and local information across modalities efficiently. Extensive experiments show that HDBFormer achieves state-of-the-art performance on the NYUDepthv2 and SUN-RGBD datasets. The code is available at: https://github.com/Weishuobin/HDBFormer.

Figures

Figures reproduced from arXiv: 2504.13579 by the authors.

**Figure 2.** The network structure diagram of the HDBFormer. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The segmentation results of HDBFormer on the NYUDepthv2 dataset [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 1.** Figure 1: Qualitative analysis of MIIM with other popular modality information [PITH_FULL_IMAGE:figures/full_fig_p006_1.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 44 canonical work pages

[1]

Attentional Feature Fusion,

Y. Dai, F. Gieseke, S. Oehmcke, Y. Wu, and K. Barnard, "Attentional Feature Fusion," in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2021, pp. 3560–3569

work page 2021
[2]

Dformer: Rethinking rgbd representation learning for semantic segmentation,

B. Yin, X. Zhang, Z. Li, L. Liu, M.-M. Cheng, and Q. Hou, "Dformer: Rethinking rgbd representation learning for semantic segmentation," arXiv preprint arXiv:2309.09668, 2023

arXiv 2023
[3]

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

J. Zhang, H. Liu, K. Yang, X. Hu, R. Liu, and R. Stiefelhagen, "CMX: Cross-modal fusion for RGB-X semantic segmentation with transformers," IEEE Transactions on Intelligent Transportation Systems, 2023

work page 2023
[4]

A survey of RGB-D image semantic segmentation by deep learning,

A. Y. Noori, "A survey of RGB-D image semantic segmentation by deep learning," in Proc. 2021 7th Int. Conf. Adv. Comput. Commun. Syst. (ICACCS), vol. 1, pp. 1953-1957, 2021

work page 2021
[5]

Non-local aggregation for RGB-D semantic segmentation,

G. Zhang, J.-H. Xue, P. Xie, S. Yang, and G. Wang, "Non-local aggregation for RGB-D semantic segmentation," IEEE Signal Processing Letters, vol. 28, pp. 658-662, 2021

work page 2021
[6]

Self-Knowledge Distillation-Based Staged Extraction and Multiview Collection Network for RGB-D Mirror Segmentation,

H. Zhang, X. Ran, and W. Zhou, "Self-Knowledge Distillation-Based Staged Extraction and Multiview Collection Network for RGB-D Mirror Segmentation," IEEE Signal Processing Letters, 2024

work page 2024
[7]

DFTI: Dual-branch Fusion Network based on Transformer and Inception for Space Non-cooperative Objects,

Z. Zhang, D. Zhou, G. Sun, Y. Hu, and R. Deng, "DFTI: Dual-branch Fusion Network based on Transformer and Inception for Space Non-cooperative Objects," IEEE Transactions on Instrumentation and Measurement, 2024

work page 2024
[8]

Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion,

Z. Zhao, H. Bai, J. Zhang, Y. Zhang, S. Xu, Z. Lin, R. Timofte, and L. Van Gool, "Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 5906-5916, 2023

work page 2023

Show all 46 references

[9]

Efficient multimodal semantic segmentation via dual-prompt learning,

S. Dong, Y. Feng, Q. Yang, Y. Huang, D. Liu, and H. Fan, "Efficient multimodal semantic segmentation via dual-prompt learning," Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024

2024
[10]

Multimodal token fusion for vision transformers,

Y. Wang, X. Chen, L. Cao, W. Huang, F. Sun, and Y. Wang, "Multimodal token fusion for vision transformers," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 12186-12195, 2022

2022
[11]

Dformer: Rethinking RGB-D representation learning for semantic segmentation,

B. Yin, X. Zhang, Z. Li, L. Liu, M.-M. Cheng, and Q. Hou, "Dformer: Rethinking RGB-D representation learning for semantic segmentation," Proceedings of the International Conference on Learning Representations (ICLR), 2024

2024
[13]

Xception: Deep learning with depthwise separable convolutions,

F. Chollet, "Xception: Deep learning with depthwise separable convolutions," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pp. 1251-1258, 2017

2017
[14]

Metaformer baselines for vision,

W. Yu, C. Si, P. Zhou, M. Luo, Y. Zhou, J. Feng, S. Yan, and X. Wang, "Metaformer baselines for vision," IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

2023
[15]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, "Attention is all you need," Proceedings of the 31st International Conference on Neural Information Processing Systems (NeurIPS), pp. 5998–6008, 2017

2017
[16]

Swin transformer: Hierarchical vision transformer using shifted windows,

Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, "Swin transformer: Hierarchical vision transformer using shifted windows," in Proc. IEEE/CVF Int. Conf. Comput. Vis., pp. 10012-10022, 2021

2021
[17]

Indoor segmentation and support inference from rgbd images,

N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, "Indoor segmentation and support inference from rgbd images," in Proc. 12th Eur. Conf. Comput. Vis. (ECCV), Florence, Italy, pp. 746-760, 2012, Springer

2012
[18]

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

S. Song, S. P. Lichtenberg, and J. Xiao, "SUN RGB-D: A RGB-D scene understanding benchmark suite," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pp. 567-576, 2015

2015
[19]

Adam: A method for stochastic optimization,

D. P. Kingma and J. Ba, "Adam: A method for stochastic optimization," Proceedings of the International Conference on Learning Representations (ICLR), 2015

2015
[20]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pp. 770-778, 2016

2016
[21]

ACNet: Attention based network to exploit complementary features for RGBD semantic segmentation,

X. Hu, K. Yang, L. Fei, and K. Wang, "ACNet: Attention based network to exploit complementary features for RGBD semantic segmentation," in Proc. IEEE Int. Conf. Image Process. (ICIP), pp. 1440-1444, 2019, IEEE

2019
[22]

Bi-directional cross-modality feature propagation with separation-and-aggregation gate for RGB-D semantic segmentation,

X. Chen, K.-Y. Lin, J. Wang, W. Wu, C. Qian, H. Li, and G. Zeng, "Bi-directional cross-modality feature propagation with separation-and-aggregation gate for RGB-D semantic segmentation," in Eur. Conf. Comput. Vis. (ECCV), pp. 561-577, 2020, Springer

2020
[23]

Deep multimodal fusion by channel exchanging,

Y. Wang, W. Huang, F. Sun, T. Xu, Y. Rong, and J. Huang, "Deep multimodal fusion by channel exchanging," Adv. Neural Inf. Process. Syst., vol. 33, pp. 4835-4845, 2020

2020
[24]

Two-stage cascaded decoder for semantic segmentation of RGB-D images,

Y. Yue, W. Zhou, J. Lei, and L. Yu, "Two-stage cascaded decoder for semantic segmentation of RGB-D images," IEEE Signal Processing Letters, vol. 28, pp. 1115-1119, 2021

2021
[25]

Delivering arbitrary-modal semantic segmentation,

J. Zhang, R. Liu, H. Shi, K. Yang, S. Rei ner, K. Peng, H. Fu, K. Wang, and R. Stiefelhagen, "Delivering arbitrary-modal semantic segmentation," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 1136-1147, 2023

2023
[26]

ComPtr: Towards Diverse Bi-source Dense Prediction Tasks via A Simple yet General Complementary Transformer,

Y. Pang, X. Zhao, L. Zhang, and H. Lu, "ComPtr: Towards Diverse Bi-source Dense Prediction Tasks via A Simple yet General Complementary Transformer," arXiv preprint arXiv:2307.12349, 2023

2023 arXiv
[27]

Omnivore: A single model for many visual modalities,

R. Girdhar, M. Singh, N. Ravi, L. Van Der Maaten, A. Joulin, and I. Misra, "Omnivore: A single model for many visual modalities," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 16102-16112, 2022

2022
[28]

MGCNet: Multilevel gated collaborative network for RGB-D semantic segmentation of indoor scene,

E. Yang, W. Zhou, X. Qian, and L. Yu, "MGCNet: Multilevel gated collaborative network for RGB-D semantic segmentation of indoor scene," IEEE Signal Processing Letters, vol. 29, pp. 2567-2571, 2022

2022
[29]

RTLNet: Recursive Triple-Path Learning Network for Scene Parsing of RGB-D Images,

Y. Yue, W. Zhou, J. Lei, and L. Yu, "RTLNet: Recursive Triple-Path Learning Network for Scene Parsing of RGB-D Images," IEEE Signal Processing Letters, vol. 29, pp. 429-433, 2021

2021
[30]

Lightweight Dual Stream Network With Knowledge Distillation for RGB-D Scene Parsing,

Y. Zhang, W. Zhou, X. Ran, and M. Fang, "Lightweight Dual Stream Network With Knowledge Distillation for RGB-D Scene Parsing," IEEE Signal Processing Letters, vol. 31, pp. 855-859, 2024

2024
[31]

Attentional feature fusion,

Y. Dai, F. Gieseke, S. Oehmcke, Y. Wu, and K. Barnard, "Attentional feature fusion," in Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis., pp. 3560-3569, 2021

2021
[32]

SegFormer: Simple and efficient design for semantic segmentation with transformers,

E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, "SegFormer: Simple and efficient design for semantic segmentation with transformers," Adv. Neural Inf. Process. Syst., vol. 34, pp. 12077-12090, 2021

2021
[33]

DCTNet: A Heterogeneous Dual-Branch Multi-Cascade Network for Infrared and Visible Image Fusion,

J. Li, L. Liu, H. Song, Y. Huang, J. Jiang, and J. Yang, "DCTNet: A Heterogeneous Dual-Branch Multi-Cascade Network for Infrared and Visible Image Fusion," IEEE Transactions on Instrumentation and Measurement, 2023

2023
[34]

MIN-MEF: Multi-scale Interaction Network for Multi-exposure Image Fusion,

Y. Yang, M. Wang, S. Huang, and W. Wan, "MIN-MEF: Multi-scale Interaction Network for Multi-exposure Image Fusion," IEEE Transactions on Instrumentation and Measurement, 2024

2024
[35]

Asymformer: Asymmetrical cross-modal representation learning for mobile platform real-time rgb-d semantic segmentation,

S. Du, W. Wang, R. Guo, R. Wang, and S. Tang, "Asymformer: Asymmetrical cross-modal representation learning for mobile platform real-time rgb-d semantic segmentation," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 7608-7615, 2024

2024
[36]

Pixel difference convolutional network for rgb-d semantic segmentation,

J. Yang, L. Bai, Y. Sun, C. Tian, M. Mao, and G. Wang, "Pixel difference convolutional network for rgb-d semantic segmentation," IEEE Transactions on Circuits and Systems for Video Technology, 2023

2023
[37]

DCANet: differential convolution attention network for RGB-D semantic segmentation,

L. Bai, J. Yang, C. Tian, Y. Sun, M. Mao, Y. Xu, and W. Xu, "DCANet: differential convolution attention network for RGB-D semantic segmentation," arXiv preprint arXiv:2210.06747, 2022

2022 arXiv
[38]

RTFormer: Efficient design for real-time semantic segmentation with transformer,

J. Wang, C. Gou, Q. Wu, H. Feng, J. Han, E. Ding, and J. Wang, "RTFormer: Efficient design for real-time semantic segmentation with transformer," Advances in Neural Information Processing Systems, vol. 35, pp. 7423--7436, 2022

2022
[39]

LKFormer: large kernel transformer for infrared image super-resolution,

F. Qin, K. Yan, C. Wang, R. Ge, Y. Peng, and K. Zhang, "LKFormer: large kernel transformer for infrared image super-resolution," Multimedia Tools and Applications, pp. 1--15, 2024

2024
[40]

Scaling up your kernels to 31x31: Revisiting large kernel design in CNNs,

X. Ding, X. Zhang, J. Han, and G. Ding, "Scaling up your kernels to 31x31: Revisiting large kernel design in CNNs," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11963--11975, 2022

2022
[41]

LKCA: Large Kernel Convolutional Attention,

C. Li, B. Zeng, Y. Lu, P. Shi, Q. Chen, J. Liu, and L. Zhu, "LKCA: Large Kernel Convolutional Attention," arXiv preprint arXiv:2401.05738, 2024

2024 arXiv
[42]

MobileNets: Efficient convolutional neural networks for mobile vision applications,

A. G. Howard, "MobileNets: Efficient convolutional neural networks for mobile vision applications," arXiv preprint arXiv:1704.04861, 2017

2017 arXiv
[43]

Xception: Deep learning with depthwise separable convolutions,

F. Chollet, "Xception: Deep learning with depthwise separable convolutions," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1251--1258, 2017

2017
[44]

GEBNet: Graph-enhancement branch network for RGB-T scene parsing,

S. Dong, W. Zhou, X. Qian, and L. Yu, "GEBNet: Graph-enhancement branch network for RGB-T scene parsing," IEEE Signal Processing Letters, vol. 29, pp. 2273--2277, 2022

2022
[45]

EM-Trans: Edge-Aware Multimodal Transformer for RGB-D Salient Object Detection,

G. Chen, Q. Wang, B. Dong, R. Ma, N. Liu, H. Fu, and Y. Xia, "EM-Trans: Edge-Aware Multimodal Transformer for RGB-D Salient Object Detection," IEEE Transactions on Neural Networks and Learning Systems, 2024

2024
[46]

Decomposed Cross-Modal Distillation for RGB-Based Temporal Action Detection,

P. Lee, T. Kim, M. Shim, D. Wee, and H. Byun, "Decomposed Cross-Modal Distillation for RGB-Based Temporal Action Detection," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 2373–2383

2023
[47]

@ @ @ @ @ @ @ @ @ @Q hTsP @ @ @ @ @ @ Ā6,B @ @ @ @ @ @ @ @ U r @ @ @ @ @ @ F8fwg #T6_m CN C ?)z?

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEco...

1999

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Attentional Feature Fusion,

Y. Dai, F. Gieseke, S. Oehmcke, Y. Wu, and K. Barnard, "Attentional Feature Fusion," in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2021, pp. 3560–3569

work page 2021

[2] [2]

Dformer: Rethinking rgbd representation learning for semantic segmentation,

B. Yin, X. Zhang, Z. Li, L. Liu, M.-M. Cheng, and Q. Hou, "Dformer: Rethinking rgbd representation learning for semantic segmentation," arXiv preprint arXiv:2309.09668, 2023

arXiv 2023

[3] [3]

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

J. Zhang, H. Liu, K. Yang, X. Hu, R. Liu, and R. Stiefelhagen, "CMX: Cross-modal fusion for RGB-X semantic segmentation with transformers," IEEE Transactions on Intelligent Transportation Systems, 2023

work page 2023

[4] [4]

A survey of RGB-D image semantic segmentation by deep learning,

A. Y. Noori, "A survey of RGB-D image semantic segmentation by deep learning," in Proc. 2021 7th Int. Conf. Adv. Comput. Commun. Syst. (ICACCS), vol. 1, pp. 1953-1957, 2021

work page 2021

[5] [5]

Non-local aggregation for RGB-D semantic segmentation,

G. Zhang, J.-H. Xue, P. Xie, S. Yang, and G. Wang, "Non-local aggregation for RGB-D semantic segmentation," IEEE Signal Processing Letters, vol. 28, pp. 658-662, 2021

work page 2021

[6] [6]

Self-Knowledge Distillation-Based Staged Extraction and Multiview Collection Network for RGB-D Mirror Segmentation,

H. Zhang, X. Ran, and W. Zhou, "Self-Knowledge Distillation-Based Staged Extraction and Multiview Collection Network for RGB-D Mirror Segmentation," IEEE Signal Processing Letters, 2024

work page 2024

[7] [7]

DFTI: Dual-branch Fusion Network based on Transformer and Inception for Space Non-cooperative Objects,

Z. Zhang, D. Zhou, G. Sun, Y. Hu, and R. Deng, "DFTI: Dual-branch Fusion Network based on Transformer and Inception for Space Non-cooperative Objects," IEEE Transactions on Instrumentation and Measurement, 2024

work page 2024

[8] [8]

Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion,

Z. Zhao, H. Bai, J. Zhang, Y. Zhang, S. Xu, Z. Lin, R. Timofte, and L. Van Gool, "Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion," in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., pp. 5906-5916, 2023

work page 2023