Pith. sign in

REVIEW 2 major objections 5 minor 51 references

GHOST: Geometry-Guided Hallucination of Opaque Surface Textures

T0 review · 2 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read A preprocessing pipeline turns transparent objects into opaque, geometry-preserving RGB so ordinary depth and 3D models work without retraining.

desk verdict Solid systems paper: a four-stage geometry-guided preprocessor that actually unlocks frozen SOTA depth/reconstruction models on transparent objects, with thorough multi-model evidence and honest ablations. read the letter →

arxiv 2607.11118 v1 pith:E5P6HQXC submitted 2026-07-13 cs.CV

Langxu Zhao , Zuan Gu , Tianhan Gao This is my paper
classification cs.CV
keywords transparentobjectsdepthestimation3Dreconstructionsurfacenormalsalphamattinggeometry-guidedsynthesisvisualfoundationmodelspreprocessing
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

Transparent surfaces break the diffuse-reflection assumptions that almost every modern depth estimator and feed-forward 3D reconstructor rely on, so those models routinely return background geometry instead of the glass surface itself. GHOST is a four-stage preprocessing stack that first isolates the transparent region, then recovers an alpha matte, a foreground intensity map and a surface-normal field, and finally synthesizes a fully opaque RGB image whose textures still obey the recovered 3D shape. Because the output is still a plain RGB photograph, any off-the-shelf monocular depth or multi-view reconstruction network can be applied unchanged. Experiments on ClearGrasp, ClearPose and related benchmarks show large gains in depth accuracy, normal angular error and reconstruction completeness for several leading foundation models. The practical claim is that one geometry-guided “opaquification” step can restore the utility of the entire existing RGB-centric 3D vision stack for transparent objects.

What carries the argument

GHOST, a four-module cascade (TransDINO for masks, TransDecomp for alpha and foreground, DAF-Net for normals, GeoSemTransNet for SPADE-conditioned texture synthesis) that turns transparent pixels into structurally consistent opaque RGB.

What would settle it

On a dataset of thick, high-refraction transparent objects where the weak-refraction assumption fails, measure whether depth and reconstruction metrics after GHOST still improve over the raw-image baseline; if they do not, the central claim collapses.

Watch

Extended reading notes

Core claim

Geometry-guided hallucination of opaque surface textures—recovering mask, alpha, foreground and normals, then synthesizing an RGB image that looks opaque yet preserves the original 3D structure—is sufficient to let unmodified state-of-the-art depth and reconstruction models produce accurate geometry on transparent objects.

Load-bearing premise

The method assumes that light transport through thin transparent objects can be treated as simple alpha blending of a foreground layer with an inpainted background, which holds only when refraction is weak.

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

2 major / 5 minor

Summary. GHOST is a four-stage preprocessing pipeline (TransDINO segmentation, TransDecomp alpha/foreground decomposition, DAF-Net surface-normal estimation, GeoSemTransNet geometry-conditioned texture synthesis) that converts transparent image regions into opaque, structure-preserving RGB so that frozen off-the-shelf depth estimators and feed-forward reconstructors can operate without retraining. The central empirical claim is that the resulting images restore photometric cues sufficiently for large, consistent gains on ClearGrasp/ClearPose (Tables 2–4) and zero-shot transfer to TROS (Table 7), with ablations (Tables 5–6) showing that partial pipelines degrade performance.

Significance. If the reported gains hold under broader conditions, the work supplies a practical, model-agnostic remedy for a long-standing failure mode of Lambertian 3D vision systems. The multi-model evaluation (DA3, MoGe2, DepthPro, Metric3Dv2, VGGT, DUSt3R, π3, MapAnything) and the demonstration that generic models plus GHOST can match or exceed specialized transparent-object estimators (Table 3) are concrete strengths. The modular design and explicit ablations make the contribution reusable and falsifiable. The principal modeling premise—the weak-refraction alpha-blending approximation—is stated rather than hidden, so the paper remains useful even where that premise is only partially valid.

major comments (2)
  1. Section 3.3 (Snell’s-law argument n1≈n2 and Eq. 7) trains TransDecomp under a weak-refraction alpha-blending model whose background prior B is produced offline by InpaintAnything. The paper notes that strong refraction yields lower alpha and that GeoSemTransNet then falls back to surface-texture synthesis, but no quantitative stratification by refractive strength, wall thickness, or multi-layer stacking is reported. Because every subsequent stage depends on the quality of α and Fg, a controlled failure-mode analysis (or an explicit statement of the operating envelope) is needed to bound the central claim that the synthesized opaque images remain geometrically faithful.
  2. Tables 2–4 and 7 report large metric improvements, yet the evaluation is confined to ClearGrasp/ClearPose (and a TROS zero-shot set) with images resized to 256×256. No multi-object occlusion, extreme lighting, or outdoor transparent-surface experiments are shown. A modest expansion of the test distribution—or an explicit limitation paragraph quantifying domain shift—would strengthen that the gains are not dataset-specific.
minor comments (5)
  1. Figure 1 caption and the surrounding text use both “GeoSemTransNet” and “GeoSem-TransNet”; a single spelling should be fixed throughout.
  2. Equation (1) and the subsequent fusion equations would benefit from an explicit statement of the feature dimensions after the DINOv3 projection, so that the cross-attention implementation is reproducible.
  3. Table 1 reports mIoU on four datasets, but the training protocol for the multi-view ClearPose subset (every 500th frame) is only briefly mentioned; a short note on whether temporal leakage was checked would improve clarity.
  4. The loss-weight choices (λsfs=10, λgrad=5, etc.) are given empirically; a one-sentence sensitivity remark or a pointer to the appendix would help readers who wish to re-implement.
  5. Several recent transparent-object depth/normal works (e.g., DKT) appear only in the comparison tables; a short related-work paragraph situating GHOST relative to diffusion-based alternatives would complete the literature framing.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; GHOST is an empirical multi-stage preprocessing pipeline whose downstream gains are measured on frozen external models and do not reduce by construction to any fitted free parameter.

full rationale

The paper’s central claim is that a four-module preprocessor (TransDINO mask extraction, TransDecomp alpha/foreground decomposition under a weak-refraction alpha-blending model, DAF-Net normal estimation, GeoSemTransNet SPADE-conditioned texture synthesis) produces opaque RGB images that improve off-the-shelf depth and reconstruction networks without retraining them. All training objectives (Dice+IoU, masked L1 reconstruction + TV, vMF NLL + edge-aware smoothness, shape-from-shading + gradient consistency) are standard supervised losses; none equate a reported metric to an input constant. Downstream tables compare frozen public models (DA3, MoGe2, DepthPro, Metric3Dv2, VGGT, DUSt3R, π3, MapAnything) on original versus GHOST-processed images; the models never see GHOST outputs during their own training. Ablations correctly demonstrate that partial pipelines degrade performance, confirming the modules are not tautological. The sole modeling premise (Snell’s-law weak-refraction approximation that justifies alpha blending) is explicitly stated as an assumption with a fallback for strong refraction; it is not used to define the evaluation metrics. No self-citation supplies a uniqueness theorem that forces the architecture, and no equation renames a fitted quantity as a “prediction.” The work is therefore self-contained against external benchmarks and exhibits only the ordinary residual circularity of any learned system that is trained on the same domain it later evaluates.

Assumptions & free parameters 5 free parameters · 4 assumptions · 4 invented entities

The central claim rests on a small set of domain modeling choices (weak refraction → alpha blending, DINOv3 structural priors, SPADE geometry injection) and a collection of hand-chosen loss weights and training hyperparameters. No new physical constants or particles are introduced; the invented entities are the four named network modules whose only evidence is the paper’s own ablations and downstream metrics.

free parameters (5)
  • λDice / λIoU for TransDINO
    Set to 0.5 / 0.5 by hand; directly controls the segmentation mask that gates every later stage.
  • λrec / λsmooth for TransDecomp
    Set to 5.0 / 0.5; balances reconstruction fidelity against foreground smoothness and therefore the quality of alpha and Fg maps.
  • λvmf / λsmooth for DAF-Net
    Weights of the von-Mises-Fisher normal loss and edge-aware smoothness; control surface-normal accuracy used as geometric prior.
  • λsfs / λgrad / λfg / λbound for GeoSemTransNet
    Set to 10.0 / 5.0 / 2.0 / 1.0; heavily prioritize shape-from-shading and gradient consistency over pixel reconstruction, shaping the hallucinated texture.
  • learning-rate schedule and image size
    Cosine decay 1e-3→1e-4, 256×256 inputs, batch 8, 70 epochs; standard but still free choices that affect final metrics.
assumptions (4)
  • domain assumption Weak refraction (n1≈n2) permits approximating transparent light transport by alpha blending of a foreground intensity and an inpainted background (Eq. 5–7).
    Stated in §3.3; supplies the only training signal for TransDecomp and therefore for all geometry and texture stages.
  • domain assumption DINOv3 dense features capture structural boundaries of transparent objects despite textural variation.
    Invoked throughout §1.2 and §3.2–3.4 as the semantic backbone for TransDINO and DAF-Net.
  • domain assumption Downstream monocular depth and feed-forward reconstruction models primarily rely on RGB photometric cues and will improve when transparent pixels are replaced by opaque, geometry-consistent textures.
    Core premise of the universal-preprocessing claim (Abstract, §1.2).
  • ad hoc to paper InpaintAnything produces a sufficiently accurate background prior B for supervised alpha/foreground decomposition.
    Used as offline pseudo-label in §3.3; quality of B is never quantified.
invented entities (4)
  • TransDINO
    purpose: Hybrid U-Net + DINOv3 segmentation that produces the transparent-object mask.
    New architecture; only evidence is Table 1 mIoU numbers on four datasets.
  • TransDecomp
    purpose: Dual-head network that disentangles alpha matte and foreground grayscale under the weak-refraction model.
    New architecture; validated only by reconstruction loss and downstream ablations.
  • DAF-Net
    purpose: Multi-modal adapter that recovers surface normals and confidence from RGB + alpha + foreground + DINO features.
    New architecture; evidence limited to normal-metric improvements inside the GHOST pipeline.
  • GeoSemTransNet
    purpose: SPADE-conditioned generator that synthesizes the final opaque textured RGB image.
    New architecture; the final hallucinated image is the sole input to all reported downstream gains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GHOST: Geometry-Guided Hallucination of Opaque Surface Textures." pith.science (2026). https://pith.science/paper/E5P6HQXC

@misc{pith2026260711118,
  author       = {Pith},
  title        = {Pith review of: GHOST: Geometry-Guided Hallucination of Opaque Surface Textures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5P6HQXC}},
  note         = {Machine review of arXiv:2607.11118}
}
read the original abstract

Transparent objects pose a fundamental challenge for depth estimation and 3D reconstruction due to their violation of Lambertian assumptions, leading to severe geometry degradation in downstream tasks. To address this, we propose a novel geometry-guided preprocessing framework \textbf{GHOST} that leverages visual foundation models to transform transparent regions into opaque, structurally consistent representations without requiring downstream model retraining. Specifically, our pipeline utilizes (1) \textbf{TransDINO} and (2) \textbf{TransDecomp} to disentangle masks and transparency physical properties, while (3) \textbf{DAF-Net} recovers surface normal priors to encode geometric curvature. Subsequently, (4) \textbf{GeoSemTransNet} integrates these multi-modal cues to synthesize a texture-rich opaque RGB image that preserves the transparent object's 3D structure. Extensive experiments demonstrate that our method significantly enhances the accuracy of state-of-the-art depth estimation and reconstruction models on transparent objects by restoring essential photometric cues.

Figures

Figures reproduced from arXiv: 2607.11118 by the authors.

Figure 1
Figure 1. Top: (1)RGB image, (2)binary mask, (3)alpha matte, (4)foreground pixels × alpha, (5)surface normal map, (6)normal-confidence map, and (7)textured RGB image.Bottom-Left: In DAF-Net, the alpha matte and coarse foreground layer serve as geometric cues. An MMSA encoder processes the input, which is then fused with DINOv3 features and progressively decoded. Two dedicated heads finally produce the surface normal map and i… view at source ↗
Figure 3
Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. Top: TransDINO takes an RGB image as input and yields a high-quality semantic mask for transparent-object segmen￾tation. Bottom: TransDecomp ingests RGB and mask and then regresses single-channel transparency and gray-scale foreground via dual heads. The result image: RGBforeground(3ch) ×alpha+ RGB255,255,255 × (1 − alpha) for visualization. data dependency. Finally, TOSQ (Ma et al., 2025) treats segmentation as a d… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Top: Qualitative results from various depth estimation models. We compare point clouds projected from raw input images versus GHOST-processed images, using predicted camera intrinsics. Bottom: Qualitative results from multiple 3D feed-forward reconstruction models, rec…
Figure 5
Figure 5. Figure 5: Visualization of segmentation results of TransDINO (Ours) and other transparent object segmentation models with high test accuracy on views from different datasets. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Visualization of 3D point cloud mapping results (via camera intrinsic parameters) for four depth estimation outputs: depth predicted by DepthAnythingV3 with GHOST preprocessing, depth directly predicted by DA3, depth predicted by DKT (a dedicated model for transparent …
Figure 7
Figure 7. Figure 7: Visualization of 3D point clouds derived from monocular depth estimation (via camera intrinsic parameters) using different models on a single real-world view, and single-view reconstruction directly performed by feedforward reconstruction models. Each method is compare…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 3 canonical work pages

  1. [1]

    and Steger, Eron , title =

    Kutulakos, Kiriakos N. and Steger, Eron , title =. International Journal of Computer Vision , year =. doi:10.1007/s11263-007-0049-9 , url =

    work page doi:10.1007/s11263-007-0049-9

  2. [2]

    Clear Grasp: 3D Shape Estimation of Transparent Objects for Manipulation , year=

    Sajjan, Shreeyak and Moore, Matthew and Pan, Mike and Nagaraja, Ganesh and Lee, Johnny and Zeng, Andy and Song, Shuran , booktitle=. Clear Grasp: 3D Shape Estimation of Transparent Objects for Manipulation , year=

  3. [3]

    2021 , eprint=

    Dex-NeRF: Using a Neural Radiance Field to Grasp Transparent Objects , author=. 2021 , eprint=

    2021

  4. [4]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Zhan, Yifan and Nobuhara, Shohei and Nishino, Ko and Zheng, Yinqiang , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2023 , pages =

    2023

  5. [5]

    2025 , eprint=

    DINOv3 , author=. 2025 , eprint=

    2025

  6. [6]

    arXiv 2023 , author=

    Foundational models defining a new era in vision: A survey and outlook. arXiv 2023 , author=. arXiv preprint arXiv:2307.13721 , year=

    arXiv 2023

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Wang, Shuzhe and Leroy, Vincent and Cabon, Yohann and Chidlovskii, Boris and Revaud, Jerome , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Wang, Jianyuan and Chen, Minghao and Karaev, Nikita and Vedaldi, Andrea and Rupprecht, Christian and Novotny, David , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    2025

Show all 51 references
  1. [9]

    2025 , eprint=

    ^3 : Permutation-Equivariant Visual Geometry Learning , author=. 2025 , eprint=

    2025

  2. [10]

    Nikhil Keetha and Norman M\"

  3. [11]

    arXiv preprint arXiv:2511.10647 , year=

    Depth Anything 3: Recovering the visual space from any views , author=. arXiv preprint arXiv:2511.10647 , year=

    arXiv

  4. [12]

    Applied Sciences , VOLUME =

    Tao, Tao and Zheng, Hong and Xiao, Jinsheng and Wu, Wenfei and Yang, Jianfeng , TITLE =. Applied Sciences , VOLUME =. 2025 , NUMBER =

    2025

  5. [13]

    arXiv preprint arXiv:2003.13948 , year=

    Segmenting Transparent Objects in the Wild , author=. arXiv preprint arXiv:2003.13948 , year=

    2003 arXiv

  6. [14]

    Cassar , keywords =

    Daniel R. Cassar , keywords =. GlassNet: A multitask deep neural network for predicting many glass properties , journal =. 2023 , issn =. doi:https://doi.org/10.1016/j.ceramint.2023.08.281 , url =

    2023 doi

  7. [15]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    Resolution-robust Large Mask Inpainting with Fourier Convolutions , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

  8. [16]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Mat: Mask-aware transformer for large hole image inpainting , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  9. [17]

    , title =

    Wang, Tianfu and Xie, Mingyang and Cai, Haoming and Shah, Sachin and Metzler, Christopher A. , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  10. [18]

    arXiv preprint arXiv:xxxx.xxxxx , year=

    DITR: Depth-Independent Transparent Object Reconstruction with Silhouette Guidance , author=. arXiv preprint arXiv:xxxx.xxxxx , year=

  11. [19]

    2023 , eprint=

    Foundational Models Defining a New Era in Vision: A Survey and Outlook , author=. 2023 , eprint=

    2023

  12. [20]

    The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Deep Depth Completion of a Single RGB-D Image , author=. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  13. [21]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Completionformer: Depth completion with convolutions and vision transformers , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  14. [22]

    RGB-D Local Implicit Function for Depth Completion of Transparent Objects , year=

    Zhu, Luyang and Mousavian, Arsalan and Xiang, Yu and Mazhar, Hammad and Eenbergen, Jozef van and Debnath, Shoubhik and Fox, Dieter , booktitle=. RGB-D Local Implicit Function for Depth Completion of Transparent Objects , year=

  15. [23]

    2019 , eprint=

    Semantic Image Synthesis with Spatially-Adaptive Normalization , author=. 2019 , eprint=

    2019

  16. [24]

    Segmenting Transparent Objects in the Wild

    Xie, Enze and Wang, Wenjia and Wang, Wenhai and Ding, Mingyu and Shen, Chunhua and Luo, Ping. Segmenting Transparent Objects in the Wild. Computer Vision -- ECCV 2020. 2020

    2020

  17. [25]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    He, Hao and Li, Xiangtai and Cheng, Guangliang and Shi, Jianping and Tong, Yunhai and Meng, Gaofeng and Prinet, Véronique and Weng, LuBin , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2021 , pages =

    2021

  18. [26]

    IEEE Transactions on Circuits and Systems for Video Technology , volume=

    Trosd: A new rgb-d dataset for transparent and reflective object segmentation in practice , author=. IEEE Transactions on Circuits and Systems for Video Technology , volume=. 2023 , publisher=

    2023

  19. [27]

    Monocular Depth Estimation and Segmentation for Transparent Object with Iterative Semantic and Geometric Fusion , year=

    Liu, Jiangyuan and Ma, Hongxuan and Guo, Yuxin and Zhao, Yuhao and Zhang, Chi and Sui, Wei and Zou, Wei , booktitle=. Monocular Depth Estimation and Segmentation for Transparent Object with Iterative Semantic and Geometric Fusion , year=

  20. [28]

    2025 , url=

    EGSA-PT:Edge-Guided Spatial Attention with Progressive Training for Monocular Depth Estimation and Segmentation of Transparent Objects , author=. 2025 , url=

    2025

  21. [29]

    arXiv preprint arXiv:2101.08461 , year=

    Segmenting transparent object in the wild with transformer , author=. arXiv preprint arXiv:2101.08461 , year=

    arXiv

  22. [30]

    IEEE Transactions on Intelligent Transportation Systems , year=

    Trans4Trans: Efficient transformer for transparent object and semantic scene segmentation in real-world navigation assistance , author=. IEEE Transactions on Intelligent Transportation Systems , year=

  23. [31]

    Sensors , VOLUME =

    Ma, Bin and Ma, Ming and Li, Ruiguang and Zheng, Jiawei and Li, Deping , TITLE =. Sensors , VOLUME =. 2025 , NUMBER =

    2025

  24. [32]

    2019 , eprint=

    ClearGrasp: 3D Shape Estimation of Transparent Objects for Manipulation , author=. 2019 , eprint=

    2019

  25. [33]

    2025 , eprint=

    HDCNet: A Hybrid Depth Completion Network for Grasping Transparent and Reflective Objects , author=. 2025 , eprint=

    2025

  26. [34]

    2024 , eprint=

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces , author=. 2024 , eprint=

    2024

  27. [35]

    European Conference on Computer Vision (ECCV) , year=

    Domain Randomization-Enhanced Depth Simulation and Restoration for Perceiving and Grasping Specular and Transparent Objects , author=. European Conference on Computer Vision (ECCV) , year=

  28. [36]

    2021 , eprint=

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows , author=. 2021 , eprint=

    2021

  29. [37]

    2021 , eprint=

    Seeing Glass: Joint Point Cloud and Depth Completion for Transparent Objects , author=. 2021 , eprint=

    2021

  30. [38]

    Frontiers in Neurorobotics , volume =

    ClueDepth Grasp: Leveraging positional clues of depth for completing depth of transparent objects , author =. Frontiers in Neurorobotics , volume =. 2022 , publisher =. doi:10.3389/fnbot.2022.1041702 , url =

    2022 doi

  31. [39]

    2023 , eprint=

    GraspNeRF: Multiview-based 6-DoF Grasp Detection for Transparent and Specular Objects Using Generalizable NeRF , author=. 2023 , eprint=

    2023

  32. [40]

    Truncated Signed Distance Function: Experiments on Voxel Size

    Werner, Diana and Al-Hamadi, Ayoub and Werner, Philipp. Truncated Signed Distance Function: Experiments on Voxel Size. Image Analysis and Recognition. 2014

    2014

  33. [41]

    NeRFrac: Neural Radiance Fields through Refractive Surface , year=

    Zhan, Yifan and Nobuhara, Shohei and Nishino, Ko and Zheng, Yinqiang , booktitle=. NeRFrac: Neural Radiance Fields through Refractive Surface , year=

  34. [42]

    2023 , eprint=

    NeRO: Neural Geometry and BRDF Reconstruction of Reflective Objects from Multiview Images , author=. 2023 , eprint=

    2023

  35. [43]

    European Conference on Computer Vision , year=

    ClearPose: Large-scale Transparent Object Dataset and Benchmark , author=. European Conference on Computer Vision , year=

  36. [44]

    TROSD: A New RGB-D Dataset for Transparent and Reflective Object Segmentation in Practice , year=

    Sun, Tianyu and Zhang, Guodong and Yang, Wenming and Xue, Jing-Hao and Wang, Guijin , journal=. TROSD: A New RGB-D Dataset for Transparent and Reflective Object Segmentation in Practice , year=

  37. [45]

    2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Deep Residual Learning for Image Recognition , author=. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    2016

  38. [46]

    Advances in Neural Information Processing Systems , pages=

    Attention is all you need , author=. Advances in Neural Information Processing Systems , pages=

  39. [47]

    2019 , eprint=

    Free-Form Image Inpainting with Gated Convolution , author=. 2019 , eprint=

    2019

  40. [48]

    2016 , eprint=

    Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs) , author=. 2016 , eprint=

    2016

  41. [49]

    2023 , eprint=

    Inpaint Anything: Segment Anything Meets Image Inpainting , author=. 2023 , eprint=

    2023

  42. [50]

    2025 , eprint=

    Diffusion Knows Transparency: Repurposing Video Diffusion for Transparent Object Depth and Normal Estimation , author=. 2025 , eprint=

    2025

  43. [51]

    2013 , publisher=

    Principles of Optics: Electromagnetic Theory of Propagation, Interference and Diffraction of Light , author=. 2013 , publisher=

    2013

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.