Pith. sign in

REVIEW 4 major objections 6 minor 32 references

DCIRNet: Depth Completion with Iterative Refinement for Dexterous Grasping of Transparent and Reflective Objects

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DCIRNet, a dual-branch RGB-depth network with cross-modal fusion and iterative refinement, raises measured dexterous-grasp success on transparent and reflective objects from 38% to 82% in the paper's real-robot experiments.

desk verdict Useful incremental depth-completion work whose headline grasp claim outruns the evidence. read the letter →

arxiv 2506.09491 v1 pith:F362FV4W submitted 2025-06-11 cs.RO cs.CV

Guanghu Xie , Zhiduo Jiang , Yonglong Zhang , Yang Liu , Zongwu Xie , Baoshi Cao , Hong Liu This is my paper
classification cs.ROcs.CV
keywords depthcompletiontransparentobjectsreflectivemultimodalfeaturefusiondexterousgraspingspatialpropagationnetworkSwinTransformerRGB-Dperception
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

DCIRNet is a depth-completion network for transparent and reflective objects, where ordinary depth sensors return holes or wrong values because light passes through or bounces off the surface. The paper's central claim is that a dual-branch architecture—one Swin Transformer (a hierarchical vision transformer) encoder for the RGB image, one for the raw depth map, linked by a learned cross-modal fusion module and an iterative spatial-propagation refinement stage—recovers the missing depth from a single RGB-D view, and that the recovered depth is good enough to change downstream behavior. On the DREDS and TransCG benchmarks the network reports best or competitive error metrics, and in real-robot grasping with the DexGraspNetV2 planner the measured success rate on ten transparent/reflective objects rises from 38% to 82% when DCIRNet's completed depth is fed to the planner. A sympathetic reader would care because transparent and reflective objects are everywhere in homes and industry, and a drop-in perception module that repairs their depth could make multi-finger manipulation substantially more reliable.

What carries the argument

The load-bearing machinery is the pair formed by the Cross-Modal Fusion Module (CMFM) and the iterative Depth Refinement module. CMFM projects each modality's features, sums them to form a fused spatial-importance map, computes per-modality softmax weights from multi-scale convolutions, cross-multiplies those weights onto the other modality's features, and then applies a depthwise convolution; this lets RGB fill holes in depth while depth constrains RGB's false edges. The refinement module is a convolutional spatial-propagation network: it predicts affinity values from decoder features, normalizes them into a convex combination over a $k\times k$ window, iterates over temporal steps, re-injects the original valid depth map under a learned confidence, and aggregates over kernel sizes $\{3,5,7\}$ and selected steps. Multi-stage supervision on both the coarse and refined predictions, with depth, normal, and gradient losses, is what drives the coarse-to-fine behavior and the reported boundary sharpness.

What would settle it

Repeat the Table III experiment with roughly 50 trials per object on the same ten objects and also substitute a second depth-completion network (for example, SwinDRNet) in place of DCIRNet; if the raw-depth baseline already succeeds most of the time, or if the 82% rate is matched by any comparable completion method, or if the 44-point gap shrinks to statistical insignificance, the paper's specific claim that DCIRNet's design drives the improvement is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery, on its own terms, is that missing depth in transparent and reflective regions can be filled well enough to change robot behavior: a single feed-forward network, given one RGB image and one raw depth map, produces completed depth that, when handed to the DexGraspNetV2 dexterous-grasp planner, raises real-world grasp success from 38% to 82% on a ten-object test set of bottles, cups, boxes, and reflective foam. Architecturally, the claim is carried by two encoder branches that each use a Swin Transformer (a hierarchical vision transformer), a cross-modal fusion module that computes spatially adaptive attention weights from multi-scale features of both modalities and exchanges them between RGB and depth, and a depth-refinement module that iteratively propagates depth values with learned affinities while blending in the valid raw-depth measurements. The network is supervised at both its coarse and refined outputs with depth, normal, and gradient losses. Benchmark results on DREDS-CatNovel and TransCG show best or comparable error metrics against the listed baselines; on DREDS-CatKnown the metrics are close to, though not uniformly better than, the SwinDRNet baseline, with the paper arguing its advantage there is sharper object boundaries.

Load-bearing premise

The load-bearing premise is that the ten-object, five-trial-per-object real-robot comparison in Table III reflects true grasp success reliably enough for the 38%-to-82% gap to be a real effect; with such a small sample, a few changed outcomes would move the headline number substantially.

Editorial extensions

If this is right

  • If DCIRNet's reported results hold, a single-view RGB-D depth-completion module is sufficient to make a dexterous grasp planner succeed on transparent and reflective objects; no multi-view capture or specialized polarization hardware is needed.
  • The completed depth can be fed to other geometry-based downstream tasks, such as object recognition and scene reconstruction, which currently degrade on transparent/reflective regions.
  • The coarse-to-fine supervision and iterative refinement recipe generalizes to novel object categories, as suggested by the DREDS-CatNovel results.
  • The grasp-success gap (38% to 82% in this experiment) implies that missing depth, rather than grasp planning itself, was a substantial cause of failure for these objects in DexGraspNetV2.
  • The reported numbers on DREDS and TransCG provide a concrete baseline for future single-view transparent/reflective depth-completion methods.

Reading between the lines

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

  • If the 38%-to-82% gain is reproducible with more trials and a broader object set, a testable prediction follows: substituting a different, equally accurate depth-completion network in the same pipeline should yield a similar grasp improvement, meaning DCIRNet's specific fusion and refinement design would matter mainly for benchmark metrics rather than for downstream success.
  • A natural next experiment would separate two possible sources of the gain: DCIRNet may help by filling transparent-region holes, or simply by denoising and regularizing the whole depth map; running the grasp pipeline on opaque objects with artificially missing depth patches would distinguish these.
  • The same front-end could be paired with a two-finger grasp planner or a mobile manipulation stack; if the gain persists there, depth completion becomes a general perception fix for transparent/reflective objects rather than a DexGraspNetV2-specific patch.
  • For real-time deployment, the paper's own stated goal of a lightweight network is the natural follow-up, since dual Swin-Transformer branches plus iterative refinement at $224\times224$ input would need to run inside a control loop; a distilled or quantized version would test whether the grasp gain survives compression.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes DCIRNet, a depth completion network for transparent and reflective objects. The architecture uses dual Swin-Transformer encoders for RGB and depth, a cross-modal fusion module, a UPerNet decoder, and an iterative spatial-propagation refinement module with multi-stage supervision. The method is evaluated on the DREDS and TransCG benchmarks using RMSE, REL, MAE, and threshold accuracy, and is integrated into the DexGraspNetV2 dexterous grasping pipeline for a real-world grasping experiment. The central claims are that DCIRNet achieves superior performance on public benchmarks and that its integration improves the grasp success rate for transparent and reflective objects by 44%, from 38% to 82%.

Significance. If the depth-completion results hold, the architectural contribution is a plausible and potentially useful system: the CMFM and depth-refinement ablations show consistent gains on DREDS-CatNovel, and the cross-dataset evaluation on TransCG is a positive sign of generalization. The paper does not ship code or trained models, so the quantitative comparisons are not independently checkable. The benchmark evidence is mixed rather than uniformly superior: DCIRNet trails SwinDRNet on DREDS-CatKnown RMSE and δ1.05, and trails TODE-Trans on TransCG RMSE and MAE. The grasping experiment is the strongest practical claim but currently lacks the controls and statistical reporting needed to attribute the 44-point gain to DCIRNet specifically. With added controls and corrected claims, the paper would be a reasonable systems contribution; in its present form the headline claim is not fully supported.

major comments (4)
  1. [IV.D.3, Table III; Abstract; V]
  2. [IV.D.1, Table I; IV.D.2, Table II]
  3. [III-B, Eqs. (4)-(5); IV.E, Table IV]
  4. [III-C, III-D, IV-C]
minor comments (6)
  1. [Fig. 3-5] The captions and image labels refer to 'MFDRNet' while the method is called DCIRNet throughout the text; please correct or explain the discrepancy.
  2. [Eq. (6)] The text states that an l1-norm constraint on κ′ ensures numerical stability, but Eq. (6) shows only a normalization by the sum of absolute values; clarify whether a separate penalty is used.
  3. [IV.C] Training is described at 224×224 input resolution, but evaluation uses 224×126 for DREDS and 240×320 for TransCG; clarify how the model is applied across resolutions.
  4. [Eq. (8)] The symbol T is used both for the set of temporal steps {0,⌊T/2⌋,T} and for the total number of iterations; use distinct names to avoid ambiguity.
  5. [Table IV] The table has no column headers or explicit row labels; please add metric names and label each configuration (e.g., single-branch baseline, +CMFM, +CMFM+DR).
  6. [Author affiliation footnote] The sentence 'All authors are with with the State Key Laboratory...' contains a duplicated 'with'; please fix.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DCIRNet is a learned depth-completion system evaluated on external public benchmarks and a physical grasping experiment, and the reported 44% improvement is a measured outcome rather than a fitted input.

full rationale

DCIRNet is a supervised depth-completion network whose outputs are completed depth maps compared against ground truth on the public DREDS and TransCG benchmarks, both external to this paper. The architecture equations (Eqs. 1-8) define a standard cross-modal fusion and spatial-propagation estimator trained with a composite loss (Eqs. 9-10); no equation defines the predicted depth in terms of the evaluation metric or the reported grasp success. The headline 44% improvement is read directly from Table III as 38.00% (19/50) versus 82.00% (41/50) observed DexGraspNetV2 successes before and after adding depth completion; this is an empirical outcome of physical trials, not a parameter fitted to that result. No load-bearing step is justified solely by a self-citation: the nearest references, SwinDRNet [11] and DexGraspNetV2 [32], are prior external systems used as baselines or tools, and the present authors do not rely on their own prior theorems to force the design. The main weaknesses--lack of a grasp-loop control using another depth-completion method, the small trial counts, and the fact that Table I shows SwinDRNet slightly better on DREDS-CatKnown RMSE and delta-1.05--are evidentiary or correctness concerns, not circularity. Therefore no circular step is identified.

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

The ledger is light on invented entities and free parameters. The main hand-chosen variables are loss weights, refinement iterations, and kernel sizes, none of which are fully reported. The paper rests on standard deep-learning assumptions about dataset reliability and pretrained backbone effectiveness, plus the domain assumption that RGB cues are informative for transparent depth.

free parameters (4)
  • Loss weights omega_D, omega_Dhat, omega_n, omega_d, omega_g
    These hand-selected weights balance the coarse/refined and depth/normal/gradient loss terms (Eqs. 9 and 10) and are never reported in Section IV-C, yet they directly affect the final performance.
  • Refinement iteration count T
    The temporal steps are defined as {0, T/2, T} in Eq. (8), but the actual value of T is not given. It controls the receptive field and computational cost of the refinement module.
  • Kernel size set K = {3,5,7}
    Kernel sizes are a design choice in Eqs. (6) and (8) and are reported, but the paper does not justify or ablate this choice.
  • Input resolution and training schedule = 224x224 (train), DREDS 224x126, TransCG 240x320 (eval)
    These hand-chosen experimental settings affect all reported metrics and are not varied or justified.
assumptions (4)
  • domain assumption Ground-truth depth in DREDS and TransCG is accurate enough to supervise training.
    The entire network is trained against these labels (Section IV-A). If the ground truth is corrupted for transparent and reflective regions, the learned mapping is biased.
  • domain assumption Swin Transformer backbones are pretrained or effectively initialized.
    Training for 20 epochs at batch size 4 on 34k to 100k images from scratch would likely underperform; the paper does not specify initialization, so it implicitly assumes standard pretrained weights.
  • standard math Equations (6)-(8) describe a stable spatial propagation scheme.
    The refinement follows CSPN++ [23]. The paper gives a slightly inconsistent definition of kappa, but the scheme is a known, stable formulation.
  • domain assumption RGB features contain complementary cues for transparent and reflective depth.
    The whole cross-modal fusion module assumes color images provide edge, texture, and context that help recover missing depth, which is plausible but not proven beyond the ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DCIRNet: Depth Completion with Iterative Refinement for Dexterous Grasping of Transparent and Reflective Objects." pith.science (2026). https://pith.science/paper/F362FV4W

@misc{pith2026250609491,
  author       = {Pith},
  title        = {Pith review of: DCIRNet: Depth Completion with Iterative Refinement for Dexterous Grasping of Transparent and Reflective Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F362FV4W}},
  note         = {Machine review of arXiv:2506.09491}
}
abstract

Transparent and reflective objects in everyday environments pose significant challenges for depth sensors due to their unique visual properties, such as specular reflections and light transmission. These characteristics often lead to incomplete or inaccurate depth estimation, which severely impacts downstream geometry-based vision tasks, including object recognition, scene reconstruction, and robotic manipulation. To address the issue of missing depth information in transparent and reflective objects, we propose DCIRNet, a novel multimodal depth completion network that effectively integrates RGB images and depth maps to enhance depth estimation quality. Our approach incorporates an innovative multimodal feature fusion module designed to extract complementary information between RGB images and incomplete depth maps. Furthermore, we introduce a multi-stage supervision and depth refinement strategy that progressively improves depth completion and effectively mitigates the issue of blurred object boundaries. We integrate our depth completion model into dexterous grasping frameworks and achieve a $44\%$ improvement in the grasp success rate for transparent and reflective objects. We conduct extensive experiments on public datasets, where DCIRNet demonstrates superior performance. The experimental results validate the effectiveness of our approach and confirm its strong generalization capability across various transparent and reflective objects.

Figures

Figures reproduced from arXiv: 2506.09491 by the authors.

Figure 1
Figure 1. The general pipeline in which a depth completion model is used [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. DCIRNet Architecture. Our network is primarily composed of an RGB encoder, a depth encoder, a multi-modal fusion module, a decoder, and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Depth Completion Visualizations of Different Models on the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Fig.5. The figure demonstrates the model’s effectiveness in [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 5
Figure 5. Figure 5: Depth Completion Visualizations of Different Models on the [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Objects in the real-world grasp experiment.1.Mineral Wa [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Real-world dexterous grasping examples. Green indicates successful [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 27 canonical work pages

  1. [1]

    Light field depth estimation for non-lambertian objects via adaptive cross operator,

    Z. Cui, H. Sheng, D. Yang, S. Wang, R. Chen, and W. Ke, “Light field depth estimation for non-lambertian objects via adaptive cross operator,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 2, pp. 1199–1211, 2023

    work page 2023

  2. [2]

    Mirror detection with the visual chirality cue,

    X. Tan, J. Lin, K. Xu, P. Chen, L. Ma, and R. W. Lau, “Mirror detection with the visual chirality cue,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 3492–3504, 2022

    work page 2022

  3. [3]

    Deep shape from polarization,

    Y . Ba, A. Gilbert, F. Wang, J. Yang, R. Chen, Y . Wang, L. Yan, B. Shi, and A. Kadambi, “Deep shape from polarization,” inComputer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIV 16. Springer, 2020, pp. 554–571

    work page 2020

  4. [4]

    Evo-nerf: Evolving nerf for sequen- tial robot grasping of transparent objects,

    J. Kerr, L. Fu, H. Huang, Y . Avigal, M. Tancik, J. Ichnowski, A. Kanazawa, and K. Goldberg, “Evo-nerf: Evolving nerf for sequen- tial robot grasping of transparent objects,” in6th annual conference on robot learning, 2022

    work page 2022

  5. [5]

    Graspnerf: Multiview-based 6-dof grasp detection for transparent and specular ob- jects using generalizable nerf,

    Q. Dai, Y . Zhu, Y . Geng, C. Ruan, J. Zhang, and H. Wang, “Graspnerf: Multiview-based 6-dof grasp detection for transparent and specular ob- jects using generalizable nerf,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 1757–1763

    work page 2023

  6. [6]

    Decomposed guided dynamic filters for efficient rgb-guided depth completion,

    Y . Wang, Y . Mao, Q. Liu, and Y . Dai, “Decomposed guided dynamic filters for efficient rgb-guided depth completion,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 2, pp. 1186–1198, 2023

    work page 2023

  7. [7]

    Dyspn: Learning dynamic affinity for image-guided depth completion,

    Y . Lin, H. Yang, T. Cheng, W. Zhou, and Z. Yin, “Dyspn: Learning dynamic affinity for image-guided depth completion,”IEEE Transac- tions on Circuits and Systems for Video Technology, vol. 34, no. 6, pp. 4596–4609, 2023

    work page 2023

  8. [8]

    Fdct: Fast depth completion for transparent objects,

    T. Li, Z. Chen, H. Liu, and C. Wang, “Fdct: Fast depth completion for transparent objects,”IEEE Robotics and Automation Letters, vol. 8, no. 9, pp. 5823–5830, 2023

    work page 2023

Show all 32 references
  1. [9]

    Diffusion-based depth inpainting for transparent and reflective objects,

    T. Sun, D. Hu, Y . Dai, and G. Wang, “Diffusion-based depth inpainting for transparent and reflective objects,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  2. [10]

    Tdcnet: Transparent objects depth completion with cnn-transformer dual- branch parallel network,

    X. Fan, C. Ye, A. Deng, X. Wu, M. Pan, and H. Yang, “Tdcnet: Transparent objects depth completion with cnn-transformer dual- branch parallel network,”arXiv preprint arXiv:2412.14961, 2024

    2024 arXiv

  3. [11]

    Domain randomization-enhanced depth simulation and restoration for perceiving and grasping specular and transparent objects,

    Q. Dai, J. Zhang, Q. Li, T. Wu, H. Dong, Z. Liu, P. Tan, and H. Wang, “Domain randomization-enhanced depth simulation and restoration for perceiving and grasping specular and transparent objects,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 374–391

    2022

  4. [12]

    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,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022

    2021

  5. [13]

    Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes,

    Q. Ha, K. Watanabe, T. Karasawa, Y . Ushiku, and T. Harada, “Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes,” in2017 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS). IEEE, 2017, pp. 5108–5115

    2017

  6. [14]

    Shapeconv: Shape-aware convolutional layer for indoor rgb-d se- mantic segmentation,

    J. Cao, H. Leng, D. Lischinski, D. Cohen-Or, C. Tu, and Y . Li, “Shapeconv: Shape-aware convolutional layer for indoor rgb-d se- mantic segmentation,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 7088–7097

    2021

  7. [15]

    Cross-modal self-attention network for referring image segmentation,

    L. Ye, M. Rochan, Z. Liu, and Y . Wang, “Cross-modal self-attention network for referring image segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 10 502–10 511

    2019

  8. [16]

    Geminifusion: Efficient pixel-wise multimodal fusion for vision trans- former,

    D. Jia, J. Guo, K. Han, H. Wu, C. Zhang, C. Xu, and X. Chen, “Geminifusion: Efficient pixel-wise multimodal fusion for vision trans- former,”arXiv preprint arXiv:2406.01210, 2024

    2024 arXiv

  9. [17]

    Cmx: Cross-modal fusion for rgb-x semantic segmentation with transform- ers,

    J. Zhang, H. Liu, K. Yang, X. Hu, R. Liu, and R. Stiefelhagen, “Cmx: Cross-modal fusion for rgb-x semantic segmentation with transform- ers,”IEEE Transactions on intelligent transportation systems, vol. 24, no. 12, pp. 14 679–14 694, 2023

    2023

  10. [18]

    Bilateral propagation network for depth completion,

    J. Tang, F.-P. Tian, B. An, J. Li, and P. Tan, “Bilateral propagation network for depth completion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 9763–9772

    2024

  11. [19]

    Learning affinity via spatial propagation networks,

    S. Liu, S. De Mello, J. Gu, G. Zhong, M.-H. Yang, and J. Kautz, “Learning affinity via spatial propagation networks,”Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  12. [20]

    Lrru: Long- short range recurrent updating networks for depth completion,

    Y . Wang, B. Li, G. Zhang, Q. Liu, T. Gao, and Y . Dai, “Lrru: Long- short range recurrent updating networks for depth completion,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 9422–9432

    2023

  13. [21]

    Non-local spatial propagation network for depth completion,

    J. Park, K. Joo, Z. Hu, C.-K. Liu, and I. So Kweon, “Non-local spatial propagation network for depth completion,” inComputer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIII 16. Springer, 2020, pp. 120–136

    2020

  14. [22]

    Learning depth with convolutional spatial propagation network,

    X. Cheng, P. Wang, and R. Yang, “Learning depth with convolutional spatial propagation network,”IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 10, pp. 2361–2379, 2019

    2019

  15. [23]

    Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion,

    X. Cheng, P. Wang, C. Guan, and R. Yang, “Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 07, 2020, pp. 10 615–10 622

    2020

  16. [24]

    Unified perceptual parsing for scene understanding,

    T. Xiao, Y . Liu, B. Zhou, Y . Jiang, and J. Sun, “Unified perceptual parsing for scene understanding,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 418–434

    2018

  17. [25]

    Dey- olo: Dual-feature-enhancement yolo for cross-modality object detec- tion,

    Y . Chen, B. Wang, X. Guo, W. Zhu, J. He, X. Liu, and J. Yuan, “Dey- olo: Dual-feature-enhancement yolo for cross-modality object detec- tion,” inInternational Conference on Pattern Recognition. Springer, 2025, pp. 236–252

    2025

  18. [26]

    Transcg: A large-scale real- world dataset for transparent object depth completion and a grasping baseline,

    H. Fang, H.-S. Fang, S. Xu, and C. Lu, “Transcg: A large-scale real- world dataset for transparent object depth completion and a grasping baseline,”IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7383–7390, 2022

    2022

  19. [27]

    Rgb-d local implicit function for depth completion of transparent objects,

    L. Zhu, A. Mousavian, Y . Xiang, H. Mazhar, J. van Eenbergen, S. Debnath, and D. Fox, “Rgb-d local implicit function for depth completion of transparent objects,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4649–4658

    2021

  20. [28]

    Clear grasp: 3d shape estimation of transparent objects for manipulation,

    S. Sajjan, M. Moore, M. Pan, G. Nagaraja, J. Lee, A. Zeng, and S. Song, “Clear grasp: 3d shape estimation of transparent objects for manipulation,” in2020 IEEE international conference on robotics and automation (ICRA). IEEE, 2020, pp. 3634–3642

    2020

  21. [29]

    Tcrnet: Transparent object depth completion with cascade refinements,

    D.-H. Zhai, S. Yu, W. Wang, Y . Guan, and Y . Xia, “Tcrnet: Transparent object depth completion with cascade refinements,”IEEE Transactions on Automation Science and Engineering, 2024

    2024

  22. [30]

    Seeing glass: joint point cloud and depth completion for transparent objects,

    H. Xu, Y . R. Wang, S. Eppel, A. Aspuru-Guzik, F. Shkurti, and A. Garg, “Seeing glass: joint point cloud and depth completion for transparent objects,”arXiv preprint arXiv:2110.00087, 2021

    2021 arXiv

  23. [31]

    Tode-trans: Transparent object depth estimation with transformer,

    K. Chen, S. Wang, B. Xia, D. Li, Z. Kan, and B. Li, “Tode-trans: Transparent object depth estimation with transformer,” in2023 IEEE international conference on robotics and automation (ICRA). IEEE, 2023, pp. 4880–4886

    2023

  24. [32]

    Dexgraspnet 2.0: Learning generative dexterous grasping in large-scale synthetic cluttered scenes,

    J. Zhang, H. Liu, D. Li, X. Yu, H. Geng, Y . Ding, J. Chen, and H. Wang, “Dexgraspnet 2.0: Learning generative dexterous grasping in large-scale synthetic cluttered scenes,” in8th Annual Conference on Robot Learning, 2024

    2024

Pith tools

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