Pith. sign in

REVIEW 5 major objections 5 minor 47 references

SR3D: Unleashing Single-view 3D Reconstruction for Transparent and Specular Object Grasping

T0 review · 5 major / 5 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 Single view rebuilds depth that glass and shiny objects hide

desk verdict A nice training-free idea for single-view transparent-object grasping, but the core scale/translation equations don't align the mesh as written and the experiments are too thin to certify the results. read the letter →

arxiv 2505.24305 v3 pith:SECBXJZP submitted 2025-05-30 cs.RO cs.CV

Mingxu Zhang , Xiaoqi Li , Jiahui Xu , Kaichen Zhou , Hojin Bae , Yan Shen , Chuyan Xiong , Hao Dong This is my paper
classification cs.ROcs.CV
keywords transparentobjectgraspingspecularsingle-view3Dreconstructiondepthmapmeshreplacementviewmatchingkeypointtraining-free
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 sets out to show that a training-free pipeline can reconstruct accurate depth for transparent and specular objects from a single RGB-D view, and that the reconstructed depth supports successful robotic grasping. The problem is real: consumer depth sensors return corrupted values on glass and shiny surfaces, so robots lack the geometry they need. SR3D sidesteps both multi-view captures and learned depth completion by generating a 3D mesh of the object from the RGB image, aligning that mesh to the reliable depth of the opaque support (the table), and then running a grasp detector on the repaired depth map. Simulation and real-robot experiments report better depth-reconstruction metrics and higher grasp success than TransCG and Asgrasp, using the same grip detector, with only one camera view. If the claim holds, robots can handle transparent and specular objects without collecting training data or setting up multiple calibrated cameras.

What carries the argument

The load-bearing mechanism is the mesh replacement module with two matching steps. View matching determines the object's 3D orientation by comparing virtual projections of the reconstructed mesh against the captured object image, using a combined score of SSIM, Laplacian edge similarity, and aspect-ratio similarity over more than 100 uniformly sampled viewpoints. Keypoint matching determines the 3D position and scale using the observation that the object's bottom contact edge with an opaque planar support is positionally consistent in 3D: three keypoints (left, center, right) sampled along the silhouette are unprojected using the table's accurate depth, the same three points are located on the mesh, and the translation and scale follow from Eqs. (6) and (8). Together they place the mesh back into the corrupted depth scene before grasp detection.

What would settle it

Take a transparent object with a curved or heavily occluded base (or rest the object on a tilted, non-planar surface), run SR3D and a trained depth-completion baseline on the same single RGB-D view, and compare each reconstruction against a painted-object ground truth; if the keypoint-derived translation or scale is visibly wrong for such objects, the central contact-edge assumption is violated and the pipeline's reported advantage should shrink or disappear.

Watch

Extended reading notes

Core claim

The central claim is that the pose and scale of a single-view reconstructed object mesh can be recovered purely from geometric and semantic cues in one RGB-D frame, making the mesh usable for manipulation. View matching renders the candidate mesh from more than 100 spherical viewpoints and picks the orientation whose image best matches the segmented object image according to a weighted score of SSIM, Laplacian-edge similarity, and aspect-ratio similarity. Keypoint matching then takes 30 evenly spaced 2D keypoints along the object's bottom contact edge, selects left, center, and right representatives, unprojects them through the accurate table depth to obtain 3D scene points, and reads the matching 3D points from the mesh; the translation is the mean of the point differences and the scale is the ratio of the pairwise distances. This mesh replacement yields a clean depth map for grasp detection. The paper argues that this alignment mechanism is what allows off-the-shelf single-view reconstruction to be transferred into a robotic grasping pipeline.

Load-bearing premise

The pipeline assumes the object's visible bottom silhouette is a contact edge resting on an opaque, near-planar support, and that the reconstructed mesh contains geometrically matching bottom points, because translation and scale are computed directly from three keypoints on that edge.

Editorial extensions

If this is right

  • Robots can grasp transparent and specular objects from a single RGB-D image, removing the need for multiple calibrated camera views or per-object training data.
  • Because the pipeline uses off-the-shelf reconstruction and grasp models, any improvement in single-view reconstruction quality should directly improve grasp success without changing the alignment logic.
  • The depth-reconstruction gains reported for both isolated objects and cluttered scenes imply the method is not limited to a clean tabletop and can be used in more realistic settings.
  • Using the same grasp detector for all compared methods isolates the benefit of the reconstructed depth, so the reported grasp-rate differences are attributed to the reconstruction rather than to the grasp model.
  • The framework works with a non-fixed camera angle, which makes it practical for mobile or hand-mounted cameras rather than requiring a rigidly calibrated observation rig.

Reading between the lines

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

  • The contact-edge assumption means the method will degrade when the object's base is occluded, rounded, or resting on a non-planar surface; a natural extension would replace the three-keypoint anchor with a more general geometric primitive (e.g., estimated support plane plus silhouette contour).
  • The orientation search is a global comparison over 100+ renderings; one could make it faster with a coarse-to-fine search or learned image descriptors while keeping the training-free property, which the paper does not explore.
  • A testable extension is to swap the reconstruction model (e.g., use a diffusion-based or Gaussian-splatting generator instead of TripoSR) and measure whether grasp success tracks mesh quality; this would quantify how much of the performance is due to the alignment mechanism versus the upstream mesh.
  • The failure analysis suggests that repositioning the object to a favorable camera angle before capture and then returning it could rescue poor reconstructions; that workaround could be formalized as an active perception policy.
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

5 major / 5 minor

Summary. SR3D proposes a training-free pipeline for grasping transparent and specular objects from a single RGB-D view. The method uses Grounded-SAM to segment the object, TripoSR to reconstruct a 3D mesh from the RGB crop, view matching over rendered views to estimate the object rotation, and keypoint matching along the object-table contact edge to estimate the translation and scale. The aligned mesh replaces the corrupted depth region, after which AnyGrasp generates grasp poses. Experiments in simulation and on a real Franka Panda arm compare SR3D against TransCG, Asgrasp, and GraspNeRF on depth-reconstruction and grasping metrics.

Significance. If the central claim holds, SR3D would be a practical result: off-the-shelf single-view reconstruction plus geometric alignment can outperform trained depth-completion and multi-view NeRF baselines without any training. The paper is clearly organized, uses publicly available components, includes a real-robot evaluation, and contains an honest failure-mode/limitations section. However, the key pose/scale alignment is internally inconsistent as written, and the empirical support is incomplete: there are no specular-object experiments, no error bars or significance tests, and the simulation baseline is reported without a full protocol. These issues are fixable within the manuscript's scope, so the result is plausible but not yet established.

major comments (5)
  1. [Section III-C2, Eqs. (6)-(8)] The similarity transform is p' = S_v R_v p + t_v, so if K^o_3D are canonical mesh coordinates, the correct translation is t_v = mean(K^s_3D - S_v R_v K^o_3D), or at least t_v = mean(K^s_3D - S_v K^o_3D) after rotation. Equation (6) computes t_v from the unscaled K^o_3D, while S_v is only introduced in Eq. (8). As written, the reconstructed mesh is misaligned by an amount proportional to (S_v - 1) times the mesh keypoint coordinates, which affects every reconstructed depth map and every grasp. The paper must state the coordinate frame of K^o_3D and the exact order of operations; if the implementation pre-scales the mesh before Eq. (6), that step needs to be documented.
  2. [Sections V-A and V-B, Tables I-III] The title and abstract claim grasping of transparent and specular objects, but the experiments only cover transparent objects. The 12-object test set is described as transparent in Fig. 3, and Table III lists only transparent objects such as bottles, goblets, phials, and beakers. No specular, glossy, or metallic object appears in the reconstruction or grasping experiments. Either add specular-object experiments or restrict the claims and title to transparent objects.
  3. [Section V-B, Table III] All grasp results are based on 10 trials per object, with no standard errors, confidence intervals, or significance tests. Several reported differences are within binomial sampling noise: SR3D and Asgrasp both achieve 8/10 on Squat Bottle, SR3D is 7/10 versus Asgrasp 8/10 on Conical Flask, and both achieve 8/10 on Phial. The claim of consistent improvement over baselines is not statistically supported. Please report more trials, confidence intervals, or a paired significance test.
  4. [Section IV] The simulation comparison with GraspNeRF is reported as a single sentence (success rate improving from 0.75 to 0.85) without a table, number of views, object split, or success criteria. In addition, the citation is inconsistent: GraspNeRF is cited as [22] in the related work but as [23] in Section IV, where [23] is Asgrasp. Provide a full protocol and results table, or remove the simulation claim.
  5. [Section V-A, Table II] In the cluttered-environment reconstruction results, SR3D's MAE (0.0607) is worse than Asgrasp's MAE (0.0380), even though SR3D has better RMSE and REL. The abstract and introduction claim SR3D outperforms all baselines; this metric reversal should be discussed explicitly, or the claim should be tempered to say SR3D is better on most metrics in most settings.
minor comments (5)
  1. [Section III-C1, Eq. (2) and Eq. (4)] The symbol W is used both for the number of SSIM patches in Eq. (2) and for the object width in Eq. (4); please use distinct symbols to avoid confusion.
  2. [Section III-C1] The view matching step is said to compare 'over 100' views, but the exact number N and the spherical sampling strategy (elevation/azimuth ranges, step size) are not specified; please report these details for reproducibility.
  3. [Section V-B] The text says reconstruction evaluations are conducted on single objects and objects in cluttered environments, but the grasping experiment in Table III lists only single objects; please clarify whether cluttered-scene grasping was performed and, if so, report the results.
  4. [Section VII] There is a typo in the limitations section: 'TripleSR' should be 'TripoSR'.
  5. [Section III-C2, Eq. (7)] Equation (7) presents T_v as an SE(3) transformation while the scale S_v is applied separately; please clarify how the final similarity transform is composed in practice, including whether the scaled mesh is rendered before translation or the translation is adjusted by the scale.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: SR3D aligns an externally reconstructed mesh to observed scene geometry, and its evaluation rests on external ground truth and baselines.

full rationale

SR3D's derivation chain is not circular. The object shape comes from the external TripoSR model; the mesh replacement module estimates rotation by 2D similarity (Eq. 5) and translation/scale by aligning contact-edge keypoints to the observed table depth (Eqs. 6-8). These are pose/scale fits from the observation, not quantities defined in terms of the output depth map. The reconstructed object depth is the transformed external mesh, so it is not merely the input depth renamed. Benchmarks are external: painted-object ground truth depth and robotic grasp success compared with TransCG, Asgrasp, and Foundation Pose. Self-citations in the related-work list (e.g., refs. [7], [8]) are not load-bearing. The paper's own limitation section (Sec. VII) attributes failures to TripoSR mesh quality and pose/scale accuracy, which also shows the output is not forced by construction. The notable Eq. (6)/(8) ordering issue is an arithmetic/implementation concern, not a circularity: it would make the alignment incorrect, but it does not make a derived quantity equivalent to its input.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The framework rests on several pretrained priors and geometric assumptions rather than on new first-principles derivation. The only quantities fitted to data are per-instance pose and scale, computed from the table contact edge; the model weights and similarity weights are chosen by hand or imported from prior work. The axioms above are the unstated conditions that would falsify the pipeline if violated.

free parameters (2)
  • Similarity weights alpha, beta, gamma = not reported
    In Eq. 5, the view matching score S = alpha*SSIM + beta*edge + gamma*ratio; the paper never states the values or how they were chosen, so they are hand-tuned parameters that could affect pose selection.
  • Number of virtual views N = over 100, not exact
    The view matching samples 'over 100' uniformly sampled camera angles on a sphere; the exact grid and rendering settings are not specified, which affects rotation accuracy.
assumptions (5)
  • domain assumption TripoSR produces a geometrically accurate 3D mesh for transparent and specular objects from a single RGB crop.
    Invoked in Section III-B; if the single-view reconstruction is inaccurate, view matching and keypoint matching propagate the error, as the authors admit in Section VII.
  • domain assumption The object rests on an opaque planar support whose depth is accurate, and the 2D contact edge between object and support corresponds to a 3D contour on the reconstructed mesh.
    Invoked in Section III-C2 for keypoint matching; the translation and scale are computed from unprojecting contact-edge keypoints with table depth, so this geometric correspondence is load-bearing.
  • ad hoc to paper The discrete set of rendered views contains a view close enough to the true object orientation for SSIM, edge, and ratio similarity to select the correct rotation.
    The paper samples 'over 100' views but gives no error bound or continuous refinement; if the true orientation is between samples, the pose has quantization error.
  • domain assumption Grounded-SAM segmentation isolates the transparent or specular object cleanly enough for TripoSR and view matching.
    Used in Section III-B; segmentation errors would contaminate the object crop and the rendered comparison.
  • domain assumption AnyGrasp can generate reliable grasps from the reconstructed depth map for transparent objects.
    Used in Section III-D; the grasp success depends on the grasp detector's behavior on mesh-replaced depth, though the paper uses the same detector for baselines.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SR3D: Unleashing Single-view 3D Reconstruction for Transparent and Specular Object Grasping." pith.science (2026). https://pith.science/paper/SECBXJZP

@misc{pith2026250524305,
  author       = {Pith},
  title        = {Pith review of: SR3D: Unleashing Single-view 3D Reconstruction for Transparent and Specular Object Grasping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SECBXJZP}},
  note         = {Machine review of arXiv:2505.24305}
}
read the original abstract

Recent advancements in 3D robotic manipulation have improved grasping of everyday objects, but transparent and specular materials remain challenging due to depth sensing limitations. While several 3D reconstruction and depth completion approaches address these challenges, they suffer from setup complexity or limited observation information utilization. To address this, leveraging the power of single view 3D object reconstruction approaches, we propose a training free framework SR3D that enables robotic grasping of transparent and specular objects from a single view observation. Specifically, given single view RGB and depth images, SR3D first uses the external visual models to generate 3D reconstructed object mesh based on RGB image. Then, the key idea is to determine the 3D object's pose and scale to accurately localize the reconstructed object back into its original depth corrupted 3D scene. Therefore, we propose view matching and keypoint matching mechanisms,which leverage both the 2D and 3D's inherent semantic and geometric information in the observation to determine the object's 3D state within the scene, thereby reconstructing an accurate 3D depth map for effective grasp detection. Experiments in both simulation and real world show the reconstruction effectiveness of SR3D.

Figures

Figures reproduced from arXiv: 2505.24305 by the authors.

Figure 1
Figure 1. The top part illustrates the pipeline of SR3D, which takes single [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The Overall Framework. The top part illustrates the overall framework flow of reconstructing depth map for transparent or specular objects, while the bottom part details the view matching and keypoint matching modules. These modules work together to determine the pose and scale required to align the reconstructed 3D object mesh with the captured 3D depth scene. Sedge = P p EI (p) · ER(p) qP p EI (p) 2 · P p ER(p) 2 … view at source ↗
Figure 3
Figure 3. Real-world test dataset. The test dataset consists of 12 transparent objects with diverse shapes. We apply white paint to their surfaces and obtain the ground truth depth maps. V. REAL-WORLD EXPERIMENTS A. Scene Reconstruction Experiments TABLE II DEPTH RECONSTRUCTION COMPARISONS ON OBJECT IN CLUTTERED ENVIRONMENT Methods RMSE↓ REL↓ MAE↓ Transcg [17] 0.1767 0.0885 0.0776 Asgrasp [23] 0.1697 0.0641 0.0380 SR3D(Ours) … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparisons of depth reconstruction results. Using the same RGB input, we visualize the reconstruction results of all methods from both the front and right views, providing a clear comparison with the ground truth. TABLE III ROBOTIC GRASPING RESULTS FOR TRANSPARENT OBJ…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 21 canonical work pages

  1. [22]

    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

    2023

  2. [23]

    Asgrasp: Generalizable transparent object reconstruction and 6-dof grasp detection from rgb-d active stereo camera,

    J. Shi, A. Yong, Y . Jin, D. Li, H. Niu, Z. Jin, and H. Wang, “Asgrasp: Generalizable transparent object reconstruction and 6-dof grasp detection from rgb-d active stereo camera,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 5441–5447

    work page 2024

  3. [1]

    Act3d: 3d feature field transformers for multi-task robotic manipulation,

    T. Gervet, Z. Xian, N. Gkanatsios, and K. Fragkiadaki, “Act3d: 3d feature field transformers for multi-task robotic manipulation,” in7th Annual Conference on Robot Learning, 2023

    work page 2023

  4. [2]

    Rvt: Robotic view transformer for 3d object manipulation,

    A. Goyal, J. Xu, Y . Guo, V . Blukis, Y .-W. Chao, and D. Fox, “Rvt: Robotic view transformer for 3d object manipulation,” inConference on Robot Learning. PMLR, 2023, pp. 694–710

    2023

  5. [3]

    Rvt- 2: Learning precise manipulation from few demonstrations,

    A. Goyal, V . Blukis, J. Xu, Y . Guo, Y .-W. Chao, and D. Fox, “Rvt- 2: Learning precise manipulation from few demonstrations,”arXiv preprint arXiv:2406.08545, 2024

    arXiv 2024

  6. [4]

    Perceiver-actor: A multi- task transformer for robotic manipulation,

    M. Shridhar, L. Manuelli, and D. Fox, “Perceiver-actor: A multi- task transformer for robotic manipulation,” inConference on Robot Learning. PMLR, 2023, pp. 785–799

    2023

  7. [5]

    3d diffusion policy,

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu, “3d diffusion policy,”arXiv preprint arXiv:2403.03954, 2024

    arXiv 2024

  8. [6]

    Peract2: Benchmarking and learning for robotic bimanual manipulation tasks,

    M. Grotz, M. Shridhar, Y .-W. Chao, T. Asfour, and D. Fox, “Peract2: Benchmarking and learning for robotic bimanual manipulation tasks,” inCoRL 2024 Workshop on Whole-body Control and Bimanual Manipulation: Applications in Humanoids and Beyond, 2024

    work page 2024

Show all 47 references
  1. [7]

    Manipllm: Embodied multimodal large language model for object-centric robotic manipulation,

    X. Li, M. Zhang, Y . Geng, H. Geng, Y . Long, Y . Shen, R. Zhang, J. Liu, and H. Dong, “Manipllm: Embodied multimodal large language model for object-centric robotic manipulation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 1...

    2024

  2. [8]

    Robomamba: Multimodal state space model for efficient robot reasoning and manipulation,

    J. Liu, M. Liu, Z. Wang, L. Lee, K. Zhou, P. An, S. Yang, R. Zhang, Y . Guo, and S. Zhang, “Robomamba: Multimodal state space model for efficient robot reasoning and manipulation,”arXiv preprint arXiv:2406.04339, 2024

    2024 arXiv

  3. [9]

    Contact- graspnet: Efficient 6-dof grasp generation in cluttered scenes,

    M. Sundermeyer, A. Mousavian, R. Triebel, and D. Fox, “Contact- graspnet: Efficient 6-dof grasp generation in cluttered scenes,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 13 438–13 444

    2021

  4. [10]

    Lift3d foundation policy: Lifting 2d large- scale pretrained models for robust 3d robotic manipulation,

    Y . Jia, J. Liu, S. Chen, C. Gu, Z. Wang, L. Luo, L. Lee, P. Wang, Z. Wang, R. Zhanget al., “Lift3d foundation policy: Lifting 2d large- scale pretrained models for robust 3d robotic manipulation,”arXiv preprint arXiv:2411.18623, 2024

    2024 arXiv

  5. [11]

    Intel realsense stereoscopic depth cameras,

    L. Keselman, J. Iselin Woodfill, A. Grunnet-Jepsen, and A. Bhowmik, “Intel realsense stereoscopic depth cameras,” inProceedings of the IEEE conference on computer vision and pattern recognition work- shops, 2017, pp. 1–10

    2017

  6. [12]

    Anygrasp: Robust and efficient grasp perception in spatial and temporal domains,

    H.-S. Fang, C. Wang, H. Fang, M. Gou, J. Liu, H. Yan, W. Liu, Y . Xie, and C. Lu, “Anygrasp: Robust and efficient grasp perception in spatial and temporal domains,”IEEE Transactions on Robotics, vol. 39, no. 5, pp. 3929–3945, 2023

    2023

  7. [13]

    Vl-grasp: a 6-dof interactive grasp policy for language-oriented objects in cluttered indoor scenes,

    Y . Lu, Y . Fan, B. Deng, F. Liu, Y . Li, and S. Wang, “Vl-grasp: a 6-dof interactive grasp policy for language-oriented objects in cluttered indoor scenes,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 976–983

    2023

  8. [14]

    Grasp anything: Combining teacher- augmented policy gradient learning with instance segmentation to grasp arbitrary objects,

    M. Mosbach and S. Behnke, “Grasp anything: Combining teacher- augmented policy gradient learning with instance segmentation to grasp arbitrary objects,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 7515–7521

    2024

  9. [15]

    An experimental assessment of depth estimation in transparent and translucent scenes for intel realsense d415, sr305 and l515,

    E. Curto and H. Araujo, “An experimental assessment of depth estimation in transparent and translucent scenes for intel realsense d415, sr305 and l515,”Sensors, vol. 22, no. 19, p. 7378, 2022

    2022

  10. [16]

    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

  11. [17]

    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

  12. [18]

    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

  13. [19]

    Grasp- ness discovery in clutters for fast and accurate grasp detection,

    C. Wang, H.-S. Fang, M. Gou, H. Fang, J. Gao, and C. Lu, “Grasp- ness discovery in clutters for fast and accurate grasp detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 15 964–15 973

    2021

  14. [20]

    V olumetric grasping network: Real-time 6 dof grasp detection in clutter,

    M. Breyer, J. J. Chung, L. Ott, R. Siegwart, and J. Nieto, “V olumetric grasping network: Real-time 6 dof grasp detection in clutter,” in Conference on Robot Learning. PMLR, 2021, pp. 1602–1611

    2021

  15. [21]

    Graspnet-1billion: A large- scale benchmark for general object grasping,

    H.-S. Fang, C. Wang, M. Gou, and C. Lu, “Graspnet-1billion: A large- scale benchmark for general object grasping,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 444–11 453

    2020

  16. [24]

    Dreamgaussian: Generative gaussian splatting for efficient 3d content creation,

    J. Tang, J. Ren, H. Zhou, Z. Liu, and G. Zeng, “Dreamgaussian: Generative gaussian splatting for efficient 3d content creation,”arXiv preprint arXiv:2309.16653, 2023

    2023 arXiv

  17. [25]

    One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization,

    M. Liu, C. Xu, H. Jin, L. Chen, M. Varma T, Z. Xu, and H. Su, “One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization,”Advances in Neural Information Processing Systems, vol. 36, pp. 22 226–22 246, 2023

    2023

  18. [26]

    One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion,

    M. Liu, R. Shi, L. Chen, Z. Zhang, C. Xu, X. Wei, H. Chen, C. Zeng, J. Gu, and H. Su, “One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 20...

    2024

  19. [27]

    Zeroshape: Regression-based zero-shot shape reconstruction,

    Z. Huang, S. Stojanov, A. Thai, V . Jampani, and J. M. Rehg, “Zeroshape: Regression-based zero-shot shape reconstruction,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10 061–10 071

    2024

  20. [28]

    Triposr: Fast 3d object reconstruction from a single image,

    D. Tochilkin, D. Pankratz, Z. Liu, Z. Huang, A. Letts, Y . Li, D. Liang, C. Laforte, V . Jampani, and Y .-P. Cao, “Triposr: Fast 3d object reconstruction from a single image,”arXiv preprint arXiv:2403.02151, 2024

    2024 arXiv

  21. [29]

    Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers,

    Z.-X. Zou, Z. Yu, Y .-C. Guo, Y . Li, D. Liang, Y .-P. Cao, and S.- H. Zhang, “Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024,...

    2024

  22. [30]

    Openlrm: Open-source large reconstruction models,

    Z. He and T. Wang, “Openlrm: Open-source large reconstruction models,” 2023

    2023

  23. [31]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Suet al., “Grounding dino: Marrying dino with grounded pre-training for open-set object detection,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 38–55

    2024

  24. [32]

    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

    2022

  25. [33]

    Said-nerf: Segmentation-aided nerf for depth comple- tion of transparent objects,

    A. Ummadisingu, J. Choi, K. Yamane, S. Masuda, N. Fukaya, and K. Takahashi, “Said-nerf: Segmentation-aided nerf for depth comple- tion of transparent objects,”arXiv preprint arXiv:2403.19607, 2024

    2024 arXiv

  26. [34]

    Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation,

    M. Hu, W. Yin, C. Zhang, Z. Cai, X. Long, H. Chen, K. Wang, G. Yu, C. Shen, and S. Shen, “Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  27. [35]

    Learning to recover 3d scene shape from a single image,

    W. Yin, J. Zhang, O. Wang, S. Niklaus, L. Mai, S. Chen, and C. Shen, “Learning to recover 3d scene shape from a single image,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 204–213

    2021

  28. [36]

    Make3d: Learning 3d scene structure from a single still image,

    A. Saxena, M. Sun, and A. Y . Ng, “Make3d: Learning 3d scene structure from a single still image,”IEEE transactions on pattern analysis and machine intelligence, vol. 31, no. 5, pp. 824–840, 2008

    2008

  29. [37]

    Robust geometry-preserving depth estimation using differ- entiable rendering,

    C. Zhang, W. Yin, G. Yu, Z. Wang, T. Chen, B. Fu, J. T. Zhou, and C. Shen, “Robust geometry-preserving depth estimation using differ- entiable rendering,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 8951–8961

    2023

  30. [38]

    Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images,

    Y . Liu, Y . Wen, S. Peng, C. Lin, X. Long, T. Komura, and W. Wang, “Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 298–315

    2022

  31. [39]

    Onepose: One-shot object pose estimation without cad models,

    J. Sun, Z. Wang, S. Zhang, X. He, H. Zhao, G. Zhang, and X. Zhou, “Onepose: One-shot object pose estimation without cad models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 6825–6834

    2022

  32. [40]

    Onepose++: Keypoint-free one-shot object pose estimation without cad models,

    X. He, J. Sun, Y . Wang, D. Huang, H. Bao, and X. Zhou, “Onepose++: Keypoint-free one-shot object pose estimation without cad models,” Advances in Neural Information Processing Systems, vol. 35, pp. 35 103–35 115, 2022

    2022

  33. [41]

    Foundationpose: Unified 6d pose estimation and tracking of novel objects,

    B. Wen, W. Yang, J. Kautz, and S. Birchfield, “Foundationpose: Unified 6d pose estimation and tracking of novel objects,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17 868–17 879

    2024

  34. [42]

    On the mathematical prop- erties of the structural similarity index,

    D. Brunet, E. R. Vrscay, and Z. Wang, “On the mathematical prop- erties of the structural similarity index,”IEEE Transactions on Image Processing, vol. 21, no. 4, pp. 1488–1499, 2011

    2011

  35. [43]

    On edge detection,

    V . Torre and T. A. Poggio, “On edge detection,”IEEE Transactions on Pattern Analysis and Machine Intelligence, no. 2, pp. 147–163, 1986

    1986

  36. [44]

    Pybullet, a python module for physics simulation for games, robotics and machine learning,

    E. Coumans and Y . Bai, “Pybullet, a python module for physics simulation for games, robotics and machine learning,” 2016

    2016

  37. [45]

    Fisher,Blender 3D Basics Beginner’s Guide: A quick and easy- to-use guide to create 3D modeling and animation using Blender 2.7

    G. Fisher,Blender 3D Basics Beginner’s Guide: A quick and easy- to-use guide to create 3D modeling and animation using Blender 2.7. Packt Publishing Ltd, 2014

    2014

  38. [46]

    Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,

    C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song, “Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,”arXiv preprint arXiv:2402.10329, 2024

    2024 arXiv

  39. [47]

    Moveit! task constructor for task-level motion planning,

    M. G ¨orner, R. Haschke, H. Ritter, and J. Zhang, “Moveit! task constructor for task-level motion planning,” in2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019, pp. 190–196

    2019

Pith tools

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