Glass Surface Segmentation with an RGB-D Camera via Weighted Feature Fusion for Service Robots

Anastasia Ioannou; Henghong Lin; Tao Wang; Yuanshui Huang; Zihan Zhu

REVIEW 5 major objections 8 minor 22 references

Glass Surface Segmentation with an RGB-D Camera via Weighted Feature Fusion for Service Robots

T0 review · 5 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a channel-wise weighted fusion of RGB and depth features is a plug-and-play module that improves glass-boundary segmentation, reporting a 7.49-point bIoU gain for PSPNet on the difficult subset of its new MJU-Glass…

desk verdict A plausible but unverified headline gain: the difficult test set is built from the compared baselines, and the dataset is the strongest contribution. read the letter →

arxiv 2508.01639 v1 pith:7MR2T4FL submitted 2025-08-03 cs.CV

Henghong Lin , Zihan Zhu , Tao Wang , Anastasia Ioannou , Yuanshui Huang This is my paper

classification cs.CV

keywords glasssurfacesegmentationRGB-DsensorfusionweightedfeaturetransparentobjectssemanticservicerobotsboundaryIoUMJU-Glassdataset

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that glass surfaces—transparent, reflective, and nearly textureless—can be segmented reliably by a small module that adaptively reweights RGB and depth features per channel and bolts onto existing segmentation networks. It also introduces MJU-Glass, a 4,900-image RGB-D dataset collected with two cameras on a wheeled service robot, so the reported gains reflect deployment-like conditions rather than curated scenes. The strongest reported effect is on boundary accuracy: adding the Weighted Feature Fusion module to PSPNet raises boundary IoU by 7.49 points on the hardest test subset. If this holds, a service robot can navigate around glass doors, windows, and partitions more safely using a standard RGB-D camera and an off-the-shelf segmentation network.

What carries the argument

The Weighted Feature Fusion module is the load-bearing mechanism. It adds the RGB feature map $r$ and the depth feature map $d$ element-wise, global-average-pools the sum to $1\times1$, passes the pooled vector through a fully connected layer that halves its dimension, then through two further fully connected layers to produce two channel-wise weight vectors. These are concatenated and passed through softmax so that $\Psi_{\mathrm{RGB}}^c + \Psi_{\mathrm{DEPTH}}^c = 1$; the weights are broadcast back to the spatial size of $r$ and $d$, and the fused map is $f = \Psi_{\mathrm{RGB}} \odot r + \Psi_{\mathrm{DEPTH}} \odot d$. The softmax normalization keeps the two modalities balanced, and the global pooling lets the module decide per channel and per image whether RGB texture or depth structure should dominate. Because the original feature maps are reused at the end, the decoder is built without skip connections, which keeps the added computation light.

What would settle it

Compare the raw depth frames with the aligned, resized frames at a glass door with known missing-depth regions; if the alignment step restores depth on the glass, the WFF depth cue is gone, and training PSPNet-WFF on depth maps with the glass holes filled should bring bIoU back down to the RGB-only level.

Watch

Extended reading notes

Core claim

The central claim is that a Weighted Feature Fusion (WFF) module, which computes per-channel softmax weights $\Psi_{\mathrm{RGB}}$ and $\Psi_{\mathrm{DEPTH}}$ with $\Psi_{\mathrm{RGB}}^c + \Psi_{\mathrm{DEPTH}}^c = 1$ and fuses backbone features as $f = \Psi_{\mathrm{RGB}} \odot r + \Psi_{\mathrm{DEPTH}} \odot d$, lets standard segmentation networks exploit the characteristic depth artifacts of glass—missing values, distorted readings, and discontinuous surfaces—while still using RGB texture when it is reliable. The paper reports that WFF improves IoU, mIoU, and boundary IoU over both RGB-only baselines and a concatenation fusion baseline on its new MJU-Glass dataset; PSPNet-WFF reaches 90.47 IoU, 93.84 mIoU, and 33.69 bIoU on the full test set, and raises bIoU by 7.49 points over RGB-only PSPNet on the difficult test set. The module is presented as plug-and-play, and the experiments with DeepLabv3+ and PSPNet on ResNet-50 support that claim.

Load-bearing premise

The depth-fusion design assumes that aligning and resizing the captured depth images preserves the missing, distorted, and discontinuous depth readings that mark glass, rather than filling or smoothing them away.

Editorial extensions

If this is right

Standard segmentation networks such as PSPNet and DeepLabv3+ can be upgraded for glass scenes by attaching WFF, without changing the backbone or the loss function.
Glass-boundary accuracy, measured by bIoU, improves more than overall mIoU, meaning the module's practical value is in avoiding collisions with edges of glass rather than in bulk region labeling.
On the difficult test set, where lighting and transparency are extreme, WFF still raises PSPNet's bIoU from 25.60 to 33.09, so the benefit persists under the conditions that matter for robot navigation.
The MJU-Glass dataset offers a 4,900-image public benchmark from an actual robot platform, so future glass-segmentation methods can be compared under deployment-like conditions.

Reading between the lines

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

A natural extension the paper does not explore is applying WFF to other two-stream inputs, such as RGB plus thermal or polarization imagery, where the softmax weighting could automatically favor the modality that is reliable in a given scene.
The dependency on raw sensor artifacts implies a stress test: on a depth camera whose driver fills missing pixels, the bIoU gains might shrink, because the glass-specific depth holes would be interpolated away before reaching the fusion module.
Because MJU-Glass records two viewpoints of the same scenes, it could additionally serve as a testbed for viewpoint sensitivity of glass-boundary cues, a question the paper leaves open.

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

Here's the short version. This paper gives service-robotics people two things worth looking at: a new RGB-D glass dataset collected from a moving robot, and a lightweight channel-wise softmax fusion module that can be grafted onto a segmentation backbone. The module is simple and clearly explained. But the headline boundary-IoU gain is computed on a difficult test set that was selected using the very baselines being compared, and no error bars are reported, so the flagship number should be treated as unverified rather than wrong.

What is genuinely new: the MJU-Glass dataset (4,900 RGB-D pairs, two-camera robot setup, public link) and the plug-and-play framing of WFF. The math is coherent; Eq. 1 is a per-channel convex combination and the softmax normalization is standard. The arithmetic checks out: 33.09 - 25.60 = 7.49 on the difficult subset. The ablation supports the main story, since the weight adjustment contributes more to boundary IoU than to mIoU. The authors cite MFFNet and Lin et al. as close prior work and explicitly position their method as simpler, which is honest.

The soft spots are substantial. The difficult test set is selected as the 400 test images with the lowest mIoU from DeepLabv3+ and PSPNet—the same baselines later used for comparison. That makes the difficult-set numbers non-independent: the module may be rewarded for matching the failure patterns of those two models, and regression to the mean can inflate the apparent gain. On the difficult set, WFF beats concatenation by only 0.88 bIoU for DeepLabv3+ but 2.92 for PSPNet, which looks like noise rather than a stable advantage. There are no multiple runs, confidence intervals, or training details, and no code. The paper never reports a comparison against Lin et al. [13], the closest prior method, so the practical-advantage claim is not established. I also noticed a table-level inconsistency: Table III gives PSPNet with F_af a bIoU of 33.60, while Table II lists the concatenation baseline at 32.66; if F_af is a different operation that is fine, but it needs to be stated. The depth-alignment concern is real but secondary: the authors should show that the missing-depth patterns used to identify glass survive resizing from 1280x800 to 640x480.

Who should read it: people working on transparent-object segmentation or RGB-D fusion for mobile robots. The dataset is the strongest contribution. For publication, the difficult-split construction should be independent of the evaluated baselines, error bars should be added, code released, and a comparison to [13] included. I recommend sending to serious peer review with the expectation of a major revision, not a desk reject.

Referee Report

5 major / 8 minor

Summary. The paper proposes a Weighted Feature Fusion (WFF) module for RGB-D glass surface segmentation, intended as a plug-and-play addition to existing segmentation backbones. The module computes per-channel softmax weights from globally pooled RGB and depth features and fuses them as f = Ψ_RGB ⊙ r + Ψ_DEPTH ⊙ d. The authors also introduce MJU-Glass, a new RGB-D dataset collected from two cameras mounted on a wheeled service robot. Experiments on DeepLabv3+ and PSPNet (both with ResNet-50) report consistent improvements in IoU, mIoU, and boundary IoU over RGB-only baselines and a simple concatenation fusion baseline, including a 7.49-point bIoU gain for PSPNet on a constructed 'difficult' test subset. The paper claims the module improves segmentation accuracy and robustness and that the dataset provides a useful benchmark for service-robot glass perception.

Significance. If the empirical claims are reliable, the WFF module would be a simple, lightweight fusion mechanism that could improve glass-boundary segmentation on standard segmentation networks, and MJU-Glass would be a valuable resource for evaluating glass segmentation from a robot perspective. The module equations are coherent, the paper ships a new publicly available dataset, and the qualitative examples are suggestive. The main significance hinges on whether the reported gains, particularly the headline bIoU improvement, are robust to evaluation setup and statistical noise, and whether the weighted mechanism itself (rather than simply adding depth features) is responsible for the improvements.

major comments (5)

[Section III-A, Table II] The 'difficult test set' is constructed by selecting the 400 test images with the lowest mIoU from DeepLabv3+ and PSPNet, which are the exact two baselines used in the comparison. This makes the difficult subset non-independent of the baselines and, because it is also used for the headline 7.49-point bIoU claim in the abstract, it can overstate the general robustness advantage of WFF. I recommend reporting the total test-set size and split, and either defining difficulty a priori (e.g., by scene properties such as lighting, transparency, or depth-missing density) or restricting the central claim to the all-set results. The all-set results still show a substantial bIoU gain for PSPNet, which mitigates the concern, but the current presentation invites a selection-bias interpretation.
[Table III vs. Table II] The ablation results contain exact numerical duplicates with the concatenation rows of Table II: the DeepLabv3+ baseline+F_af row (89.76/93.36/39.91) is identical to the DeepLabv3+-concat row, and the PSPNet baseline+F_af IoU and mIoU (90.21/93.67) match PSPNet-concat exactly while its bIoU is 33.60 versus 32.66. This suggests a copy/paste or tabulation error, and it also reveals that the adaptive weighting component F_aw adds only 0.09 bIoU for PSPNet and 0.44 bIoU for DeepLabv3+ on the all set. These numbers contradict the text's assertion that F_aw 'plays a more crucial role in refining the boundaries.' Please reconcile the tables and either correct the errors or revise the interpretation. If F_aw provides only marginal benefits over simple feature summation, the paper's central technical claim about adaptive weighting needs to be substantially softened.
[Section III-C] All experiments appear to be from a single training run, with no standard deviations, confidence intervals, or multiple seeds. Several reported gains are small relative to typical segmentation variance, e.g., the DeepLabv3+ all-set mIoU gain of 0.44 points (93.17 to 93.61) and the difficult-set mIoU gain of 0.24 points (92.12 to 92.36). Without error bars, the reader cannot tell whether the WFF advantage over concatenation is meaningful or within noise. Please report mean and standard deviation over at least three seeds, or provide a statistical significance test.
[Section II-A, Section III-C] The claim that the module is a 'plug-and-play' solution for various deep neural network backbones is supported by experiments on only two backbones, both using the same ResNet-50 encoder. The introduction and Fig. 1 also show UNet and SegViT qualitatively, but no quantitative results are reported for them. Please add quantitative results with at least one additional backbone family, or explicitly limit the scope of the plug-and-play claim to the two tested architectures.
[Section III-C] The experiments compare only against RGB-only baselines and a simple concatenation fusion baseline; there are no comparisons with existing glass-specific RGB-D methods such as Lin et al. [13] (cross-modal context mining) or Zhu et al. [12] (MFFNet), both of which are cited in the introduction. Without such comparisons, it is unclear whether the proposed module is competitive with current state-of-the-art glass detection methods. Please add comparisons with these or similar methods on MJU-Glass, or on existing benchmarks such as RGB-D GSD, to support the claim of broad usefulness.

minor comments (8)

[Section III-A] The dataset URL 'https://github.com/weduake/MJU Glass' contains a space; the correct URL should likely be 'https://github.com/weduake/MJU-Glass'.
[Abstract and Section III-C] The phrase '7.49% improvement in bIoU' refers to an absolute percentage-point increase from 25.60 to 33.09. Please clarify whether the reported number is an absolute or relative change to avoid ambiguity.
[Section II-C, Eq. (1)] There is a typesetting error in the definition of Ψ_DEPTH: 'RC a' should be 'R^C'. Additionally, the channel-index notation is used inconsistently: the sum constraint uses superscript c while the weights are defined as vectors; please define Ψ_c^RGB and Ψ_c^DEPTH clearly for each channel c.
[Section II-C] The text says that both shallow and deep features are fused, but it is not clear from Fig. 3 and the surrounding equations whether the WFF module is applied separately at two feature scales or only once. Please specify how the module is instantiated for the shallow and deep branches, and describe where in Fig. 2 each WFF operation occurs.
[Section III-B] The dataset split is not described: the total number of images is given as 4,900, but the numbers of training, validation, and test images are not reported. Please also state whether depth missing values were used as-is or preprocessed (e.g., interpolated) before resizing and alignment.
[Section III-C, Eq. (8)] The sentence accompanying Eq. (8) says 'FP_Glass and FN_Glass represent true positives, false positives, and true negatives' — it should say 'false negatives.' This is a typos in the definition.
[Section II-C, Eq. (7)] In the softmax equation, the vectors Ψ_1 and Ψ_2 are not shown with channel indices, although they are per-channel. Adding explicit subscripts (e.g., Ψ_1^c) would make the normalization over the concatenated channel axis easier to follow.
[Section I, Fig. 1] The teaser figure shows qualitative improvements for UNet, PSPNet, and SegViT, but only PSPNet and DeepLabv3+ are evaluated quantitatively. Either include quantitative results for UNet and SegViT or remove them from the teaser to avoid overclaiming.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; WFF is an empirical learned-fusion method evaluated on held-out data.

full rationale

No significant circularity. The WFF module's fusion weights are learned parameters (Eqs. 4-7) trained by binary cross-entropy on a training split and evaluated on held-out test images; there is no equation in which the predicted quantity appears as an input or fitted target. The paper makes no first-principles derivation claim and invokes no uniqueness theorem. Prior citations to the authors' earlier work, such as [9], are contextual and are not load-bearing for the WFF design or the reported gains. The only notable concern is the construction of the 'difficult test set' as the 400 images with lowest mIoU from DeepLabv3+ and PSPNet, which introduces selection bias and may inflate the headline 7.49 bIoU gain relative to an independently defined hard split; however, this is an evaluation-validity issue, not circular reasoning, because the WFF weights are not fitted to that subset and the improvement is not true by construction. The central claim is empirically testable and self-contained against standard baselines and a released dataset.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's empirical claim rests on the domain assumption that depth artifacts reliably mark glass, on the backbone's feature quality, and on a model-dependent difficult test set. Training details and learned-weight statistics are unreported. No physical entities are invented.

free parameters (3)

WFF fusion weights Psi_RGB and Psi_DEPTH = learned, not reported
Per-channel softmax weights are fit during training; their values and behavior across lighting conditions are not reported.
Dimensions of FC1, FC21, FC22 in the WFF module = unspecified beyond halving then restoring channel dimension
The paper says FC1 halves the feature dimension and the second two FC layers restore it (Eqn. 6), but exact sizes are needed for reproduction.
Training hyperparameters (optimizer, learning rate, epochs, batch size, augmentation, split) = not reported
Sec. III-B gives no training details; the reported numbers cannot be regenerated without these choices.

assumptions (4)

domain assumption Glass surfaces produce reliable anomalous depth patterns (missing values, distortions, discontinuities) with the Orbbec Gemini Pro RGB-D camera, and these patterns survive alignment and resizing to 640x480.
Invoked in Sec. I and Sec. III-A; the whole WFF approach depends on depth carrying a usable glass cue.
domain assumption A ResNet-50 backbone pretrained on ImageNet, trained on MJU-Glass, yields separable RGB and depth features whose per-channel weighted sum improves segmentation.
Secs. II-A and II-B assume the backbone features are complementary and that global channel statistics suffice for fusion.
ad hoc to paper The difficult test set, defined as the 400 images with lowest mIoU from DeepLabv3+ and PSPNet, is a valid robustness benchmark rather than a biased subset.
Sec. III-A introduces this subset; its model-dependent construction is not justified as representative of hard conditions.
standard math Softmax normalization in Eqn. 7 produces valid channel weights with Psi_RGB + Psi_DEPTH = 1.
Standard property of softmax; used to define the fusion weights.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Glass Surface Segmentation with an RGB-D Camera via Weighted Feature Fusion for Service Robots." pith.science (2026). https://pith.science/paper/7MR2T4FL

@misc{pith2026250801639,
  author       = {Pith},
  title        = {Pith review of: Glass Surface Segmentation with an RGB-D Camera via Weighted Feature Fusion for Service Robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7MR2T4FL}},
  note         = {Machine review of arXiv:2508.01639}
}

read the original abstract

We address the problem of glass surface segmentation with an RGB-D camera, with a focus on effectively fusing RGB and depth information. To this end, we propose a Weighted Feature Fusion (WFF) module that dynamically and adaptively combines RGB and depth features to tackle issues such as transparency, reflections, and occlusions. This module can be seamlessly integrated with various deep neural network backbones as a plug-and-play solution. Additionally, we introduce the MJU-Glass dataset, a comprehensive RGB-D dataset collected by a service robot navigating real-world environments, providing a valuable benchmark for evaluating segmentation models. Experimental results show significant improvements in segmentation accuracy and robustness, with the WFF module enhancing performance in both mean Intersection over Union (mIoU) and boundary IoU (bIoU), achieving a 7.49% improvement in bIoU when integrated with PSPNet. The proposed module and dataset provide a robust framework for advancing glass surface segmentation in robotics and reducing the risk of collisions with glass objects.

Figures

Figures reproduced from arXiv: 2508.01639 by the authors.

**Figure 2.** Overall model architecture. In this model, features of glass are fused [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** An illustration of the proposed Weighted Feature Fusion (WFF) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Examples RGB-D images from the MJU-Glass dataset, highlighting [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 6.** Figure 6: Example RGB-D glass surface segmentation results using DeepLabv3+ [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages

[13]

Leveraging rgb-d data with cross-modal context mining for glass surface detection,

J. Lin, Y .-H. Yeung, S. Ye, and R. W. Lau, “Leveraging rgb-d data with cross-modal context mining for glass surface detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 5, 2025, pp. 5254–5261

work page 2025
[12]

Mffnet: Multimodal feature fusion network for rgb-d transparent object detection,

L. Zhu, T. Li, Y . Ning, and Y . Zhang, “Mffnet: Multimodal feature fusion network for rgb-d transparent object detection,”International Journal of Advanced Robotic Systems, vol. 21, no. 5, p. 17298806241283373, 2024

work page 2024
[1]

U-net: Convolutional networks for biomedical image segmentation,

O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,”ArXiv, vol. abs/1505.04597, 2015. [Online]. Available: https://api.semanticscholar.org/CorpusID:3719281

arXiv 2015
[2]

Pyramid scene parsing network,

H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing network,”2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6230–6239, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:5299559

work page 2017
[3]

Segvit: Semantic segmentation with plain vision transformers,

B. Zhang, Z. Tian, Q. Tang, X. Chu, X. Wei, C. Shen, and Y . Liu, “Segvit: Semantic segmentation with plain vision transformers,” ArXiv, vol. abs/2210.05844, 2022. [Online]. Available: https://api. semanticscholar.org/CorpusID:252846611

arXiv 2022
[4]

Fully convolutional networks for semantic segmentation,

E. Shelhamer, J. Long, and T. Darrell, “Fully convolutional networks for semantic segmentation,”2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3431–3440, 2014. [Online]. Available: https://api.semanticscholar.org/CorpusID:1629541

work page 2015
[5]

Rethinking atrous convolution for semantic image segmentation,

L.-C. Chen, G. Papandreou, F. Schroff, and H. Adam, “Rethinking atrous convolution for semantic image segmentation,”ArXiv, vol. abs/1706.05587, 2017. [Online]. Available: https://api.semanticscholar. org/CorpusID:22655199

arXiv 2017
[6]

Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers,

S. Zheng, J. Lu, H. Zhao, X. Zhu, Z. Luo, Y . Wang, Y . Fu, J. Feng, T. Xiang, P. H. S. Torr, and L. Zhang, “Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers,”2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6877–6886, 2020. [Online]. Available: https://api.semanticscholar.org/Corpus...

work page 2021

Show all 22 references

[7]

Glass segmentation with rgb-thermal image pairs,

D. Huo, J. Wang, Y . Qian, and Y .-H. Yang, “Glass segmentation with rgb-thermal image pairs,”IEEE Transactions on Image Processing, vol. 32, pp. 1911–1926, 2022. [Online]. Available: https://api. semanticscholar.org/CorpusID:248118556

1911
[8]

Glass segmentation using intensity and spectral polarization cues,

H. Mei, B. Dong, W. Dong, J. Yang, S.-H. Baek, F. Heide, P. Peers, X. Wei, and X. Yang, “Glass segmentation using intensity and spectral polarization cues,”2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12 612–12 621, 2022. [Online]. Available:...

2022
[9]

Glass object localization by joint inference of boundary and depth,

T. Wang, X. He, and N. Barnes, “Glass object localization by joint inference of boundary and depth,”Proceedings of the 21st International Conference on Pattern Recognition (ICPR2012), pp. 3783–3786, 2012. [Online]. Available: https://api.semanticscholar.org/CorpusID:18189708

2012
[10]

Glass detection and recognition based on the fusion of ultrasonic sensor and rgb-d sensor for the visually impaired,

Z. Huang, K. Wang, K. Yang, R. Cheng, and J. Bai, “Glass detection and recognition based on the fusion of ultrasonic sensor and rgb-d sensor for the visually impaired,” inSecurity + Defence, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:70281006

2018
[11]

Glass detection in simultaneous localization and mapping of mobile robot based on rgb-d camera,

Y . Zhao, H. Li, S. Jiang, H. Li, Z. Zhang, and H. Zhu, “Glass detection in simultaneous localization and mapping of mobile robot based on rgb-d camera,”2023 IEEE International Conference on Mechatronics and Automation (ICMA), pp. 549–556, 2023. [Online]. Available: https://ap...

2023
[14]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,”2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2015. [Online]. Available: https://api.semanticscholar.org/CorpusID:206594692

2016
[15]

Depth-aware mirror segmentation,

H. Mei, B. Dong, W. Dong, P. Peers, X. Yang, Q. Zhang, and X. Wei, “Depth-aware mirror segmentation,”2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3043–3052,

2021
[16]

Matterport3d: Learning from rgb-d data in indoor environments,

A. X. Chang, A. Dai, T. A. Funkhouser, M. Halber, M. Nießner, M. Savva, S. Song, A. Zeng, and Y . Zhang, “Matterport3d: Learning from rgb-d data in indoor environments,”2017 International Conference on 3D Vision (3DV), pp. 667–676, 2017. [Online]. Available: https://api.semant...

2017
[17]

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,”2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 567–576, 2015. [Online]. Available: https://api.semanticscholar.org/CorpusID:6242669

2015
[18]

Rethinking rgb-d salient object detection: Models, data sets, and large-scale benchmarks,

D.-P. Fan, Z. Lin, J. Zhao, Y . Liu, Z. Zhang, Q. Hou, M. Zhu, and M.-M. Cheng, “Rethinking rgb-d salient object detection: Models, data sets, and large-scale benchmarks,”IEEE Transactions on Neural Networks and Learning Systems, vol. 32, pp. 2075–2089, 2019. [Online]. Availab...

2019
[19]

Joint 2d-3d-semantic data for indoor scene understanding,

I. Armeni, S. Sax, A. Zamir, and S. Savarese, “Joint 2d-3d-semantic data for indoor scene understanding,”ArXiv, vol. abs/1702.01105, 2017. [Online]. Available: https://api.semanticscholar.org/CorpusID:2730848

2017 arXiv
[20]

Monocular depth estimation for glass walls with context: A new dataset and method,

Y . Liang, B. Deng, W. Liu, J. Qin, and S. He, “Monocular depth estimation for glass walls with context: A new dataset and method,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, pp. 15 081–15 097, 2023. [Online]. Available: https://api.semanticscholar...

2023
[21]

Encoder-decoder with atrous separable convolution for semantic image segmentation,

L.-C. Chen, Y . Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder-decoder with atrous separable convolution for semantic image segmentation,” inEuropean Conference on Computer Vision, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:3638670

2018
[2021]

Available: https://api.semanticscholar.org/CorpusID: 235357151

[Online]. Available: https://api.semanticscholar.org/CorpusID: 235357151

Pith tools

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

[1] [13]

Leveraging rgb-d data with cross-modal context mining for glass surface detection,

J. Lin, Y .-H. Yeung, S. Ye, and R. W. Lau, “Leveraging rgb-d data with cross-modal context mining for glass surface detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 5, 2025, pp. 5254–5261

work page 2025

[2] [12]

Mffnet: Multimodal feature fusion network for rgb-d transparent object detection,

L. Zhu, T. Li, Y . Ning, and Y . Zhang, “Mffnet: Multimodal feature fusion network for rgb-d transparent object detection,”International Journal of Advanced Robotic Systems, vol. 21, no. 5, p. 17298806241283373, 2024

work page 2024

[3] [1]

U-net: Convolutional networks for biomedical image segmentation,

O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,”ArXiv, vol. abs/1505.04597, 2015. [Online]. Available: https://api.semanticscholar.org/CorpusID:3719281

arXiv 2015

[4] [2]

Pyramid scene parsing network,

H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing network,”2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6230–6239, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:5299559

work page 2017

[5] [3]

Segvit: Semantic segmentation with plain vision transformers,

B. Zhang, Z. Tian, Q. Tang, X. Chu, X. Wei, C. Shen, and Y . Liu, “Segvit: Semantic segmentation with plain vision transformers,” ArXiv, vol. abs/2210.05844, 2022. [Online]. Available: https://api. semanticscholar.org/CorpusID:252846611

arXiv 2022

[6] [4]

Fully convolutional networks for semantic segmentation,

E. Shelhamer, J. Long, and T. Darrell, “Fully convolutional networks for semantic segmentation,”2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3431–3440, 2014. [Online]. Available: https://api.semanticscholar.org/CorpusID:1629541

work page 2015

[7] [5]

Rethinking atrous convolution for semantic image segmentation,

L.-C. Chen, G. Papandreou, F. Schroff, and H. Adam, “Rethinking atrous convolution for semantic image segmentation,”ArXiv, vol. abs/1706.05587, 2017. [Online]. Available: https://api.semanticscholar. org/CorpusID:22655199

arXiv 2017

[8] [6]

Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers,

S. Zheng, J. Lu, H. Zhao, X. Zhu, Z. Luo, Y . Wang, Y . Fu, J. Feng, T. Xiang, P. H. S. Torr, and L. Zhang, “Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers,”2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6877–6886, 2020. [Online]. Available: https://api.semanticscholar.org/Corpus...

work page 2021