From Transparent Labware Segmentation to Collision Avoidance: A Real-Time Edge-Aware Perception Pipeline

Chen Qian; Junlin Xiong; Shijun Ding; Weiwei Shang

REVIEW 4 major objections 5 minor 27 references

From Transparent Labware Segmentation to Collision Avoidance: A Real-Time Edge-Aware Perception Pipeline

T0 review · 4 major / 5 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 This paper argues that explicit edge-aware instance segmentation, combined with multi-view centroid triangulation and conservative bounding-volume modeling, is a practical perception-to-action solution for robot collision avoidance in…

desk verdict A genuinely useful dataset and a solid real-time segmentation method, but the collision-avoidance headline leans on an unreported hand-selected safety margin that a reviewer must force into the open. read the letter →

arxiv 2608.04769 v1 pith:DNGYDF5S submitted 2026-08-05 cs.RO

Shijun Ding , Chen Qian , Weiwei Shang , Junlin Xiong This is my paper

classification cs.RO

keywords transparentobjectsegmentationinstanceedge-awareperceptionrobotcollisionavoidancemulti-view3Destimationlaboratoryglasswaredatasetreal-timeinferenceboundaryF-score

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper sets out to make transparent laboratory glassware visible to robot motion planning using only RGB cameras. Transparent objects defeat ordinary segmentation because refraction, specular highlights, and missing interior texture leave few stable cues, but their outer contours remain reliable. The authors build a one-stage instance segmentation network that adds an explicit edge-detection branch, edge-guided attention fusion, and a parameter-free attention module to a YOLOv5-Seg backbone, and train it on LabGlass-IS, a new 3,485-image, 21-category dataset of real labware. They then back-project each mask's centroid through multiple camera views, triangulate a 3D position, and wrap each vessel in a conservatively sized cube for collision checking. On their benchmark the model reaches a boundary F-score of 97.80 and an mAP 50:95 of 82.2 at 7.1 ms per frame, and a real robot using the full loop avoids collisions in 14 of 15 trials.

What carries the argument

The load-bearing components are: (1) a multi-scale edge branch that predicts an edge confidence map under explicit supervision and feeds edge-aware features into the neck through a Bottleneck Attention Module; (2) a parameter-free SimAM attention module that reweights neurons by an energy-based separability measure; (3) multi-view least-squares triangulation of 2D mask centroids to estimate each vessel's 3D centroid; and (4) an axis-aligned cubic bounding volume, centered at the estimated centroid, whose size is conservatively selected by hand to cover the object plus segmentation uncertainty. The 2D geometric centroid of the predicted mask, back-projected along viewing rays, is the key anchor that ties perception to planning without dense 3D reconstruction.

What would settle it

Place a slender transparent glass rod inside a wider flask so refraction displaces the rod's mask centroid; if the triangulated cube centered on that centroid fails to enclose the rod, the robot will contact it, directly testing whether the hand-selected cube size covers the worst perceived centroid bias for slender objects.

Watch

Extended reading notes

Core claim

The central claim is that for transparent vessels the boundary contour is the dependable perceptual signal, and an architecture that explicitly extracts, supervises, and re-injects edge information can produce real-time instance masks accurate enough to anchor 3D obstacle avoidance. The paper reports that the edge-augmented YOLOv5-Seg model outperforms all compared general-purpose segmenters on boundary quality (BF 97.80 versus 78.87 for YOLO-prompted FastSAM and 88.91 for PointRend), while running at 7.1 ms per frame with 17.18 M parameters, and that multi-view triangulation of mask centroids gives an average 3D position error of 38.0 mm, which the conservative cubes are sized to absorb. The real-robot experiments report 93.3% collision-avoidance success across 15 trials, supporting the paper's conclusion that explicit edge awareness plus conservative geometric modeling provides a practical perception-to-action solution in cluttered transparent-object scenes.

Load-bearing premise

The entire safety margin rests on the assumption that a mask centroid triangulated from a few views lands close enough to the true vessel center that a manually chosen cube size around it fully covers the object, and that cube size is selected by hand rather than derived from the measured 38.0 mm centroid error.

Editorial extensions

If this is right

The edge-augmented model can segment transparent labware at 7.1 ms per frame with 17.18 M parameters, making closed-loop collision checking between replans practical on robot hardware.
Multi-view mask-centroid triangulation yields 3D positions with an average error of 38.0 mm, and the conservative cubes are sized to absorb this error, so the robot can avoid fragile vessels without dense 3D reconstruction.
The Boundary F-score gain is largest on slender objects (98.21 vs 88.73 for YOLACT), so the edge branch specifically recovers thin, refractive structures that other segmenters miss.
LabGlass-IS gives the community a real-world benchmark with per-instance masks for 21 transparent vessel categories, supporting training and evaluation of future perception pipelines for lab automation.

Reading between the lines

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

The same edge-cue principle should transfer to other transparent or specular objects beyond labware, such as glass panels, vials, or windshields, since boundary contours remain the stable cue wherever interior texture is absent; the paper only claims labware.
The conservative-cube strategy trades workspace for safety: with a measured maximum centroid error of 60.5 mm, the cube margin must grow in cluttered scenes, which could eventually block feasible paths; a per-category margin derived from the error distribution would free workspace.
Because the reported 38.0 mm centroid error exceeds typical end-effector positioning accuracy, the pipeline favors collision avoidance over manipulation; extending it to grasping would require a finer pose estimate than a cube around the centroid.

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

The thing you should know: this is a solid applied vision paper, not a breakthrough. The real contribution is LabGlass-IS, a 3,485-image, 21-class, 6,099-instance dataset of real laboratory glassware with pixel-level masks, plus a lightweight edge-aware instance segmentation model that beats existing baselines on boundary quality (Boundary F-score 97.80) at 7.1 ms per frame. That part is credible. The ablations are internally consistent, the comparisons are fair, and the code and video are public. The paper deserves a serious referee, mainly because the dataset is useful and the segmentation results are likely reproducible.

The soft spots are real but, for the most part, not fatal. The stress-test concern lands on reading: the collision-avoidance claim rests on a bounding cube whose size is "conservatively selected" (Section III-C) without stating dimensions, margins, or how the margin relates to the measured centroid errors (mean 38.0 mm, max 60.5 mm). The text says the volumes are sized to enclose the objects "with this margin," but no numbers appear. So the 93.3% avoidance rate (14/15) cannot be independently assessed. With only 15 trials, the confidence interval is wide, and the single failure may reflect either an undersized or oversized cube. This does not sink the paper, but a reviewer should demand the missing geometry and, ideally, more trials.

There are two more minor issues. The dataset itself is not stated to be released, only the code and video; since the dataset is a major claimed contribution, the release status needs to be explicit. And the segmentation metrics come from single runs, with no error bars or multi-seed variance, so the small deltas between ablations should be read cautiously.

On the positive side, the method is not a pile of tricks: the edge branch gives a large boundary improvement, SimAM contributes modestly, and the system integration from mask to 3D centroid to conservative obstacle is a sensible, practical pipeline. The paper is honest in its conclusion that this is for collision avoidance, not fine manipulation.

My take: send it to peer review. The dataset alone justifies referee time. The author-revision should force quantification of the safety margin, explicit dataset release or a clear reason not to, and ideally a few more robot trials. This is a useful paper that would be stronger with those numbers in the open.

Referee Report

4 major / 5 minor

Summary. This paper presents a perception-to-action pipeline for transparent laboratory glassware: an edge-aware instance segmentation network built on YOLOv5-Seg with an auxiliary edge branch, edge-guided attention fusion, and a parameter-free SimAM module; a new real-world dataset, LabGlass-IS, containing 3,485 images, 21 categories, and 6,099 instance annotations; and a collision-avoidance module that triangulates mask centroids across multiple views and constructs conservative axis-aligned cubic obstacle volumes. The experiments report a Boundary F-score of 97.80, mAP50:95 of 82.2, and 7.1 ms inference on the segmentation side, and a 93.3% collision-avoidance success rate over 15 real-robot trials.

Significance. The paper's strengths are a carefully constructed real-world dataset with instance-level annotations, a lightweight architecture with a small parameter footprint, and a complete robot demonstration with code and video links. The evaluation is empirical and does not rely on circular reasoning or fitted-parameter derivations. If the reported numbers hold, the work would be a practically useful demonstration that low-cost edge-aware segmentation plus coarse geometric modeling can support real-time collision avoidance for transparent objects. However, the central robot-safety claim rests on an unquantified hand-selected cube margin and a small number of trials, and the segmentation improvements are reported without variance. These issues need to be addressed before the claims are fully supported.

major comments (4)

[III-C, Table IV] The safety margin of the conservative bounding cube is never quantified. The text states only that "The cube size is conservatively selected to cover the spatial extent of the object while accounting for segmentation uncertainty and geometric variations" (Section III-C). Since the measured centroid errors are mean 38.0 mm and max 60.5 mm (Table III), the cube side length and the margin used determine whether the obstacle volume actually contains the physical glassware; a margin that is too small makes contact possible, while one that is too large can make collision-free paths infeasible. The 93.3% success rate in Table IV is therefore conditional on an unstated hand-selected parameter. Please report the cube dimensions and margins used for each object category or scene, and provide a sensitivity analysis of success rate and planning feasibility as the margin is varied.
[Table IV] The collision-avoidance evaluation uses only 15 trials total (14/15 successes). The resulting 95% confidence interval for the true success rate is approximately 68%–99.8%, so the headline 93.3% is not statistically strong. Please increase the number of trials, particularly in the Hard condition, report confidence intervals, and describe the failure mode of the single failed trial.
[Tables I and II] All segmentation metrics are single-run values without standard deviations or multiple seeds. The key reported improvements in the ablation are modest (BF 96.67→97.80; mAP50:95 81.3→82.2), and the comparison with PointRend shows a higher mAP50:95 for PointRend (84.35) while the proposed method wins on BF. Without variance estimates, it is unclear whether these differences are significant. Please report mean±std over at least three training runs (or cross-validation) and, where appropriate, paired tests for BF and mAP.
[Table I / Related Work] The related work discusses LBSNet and Trans2Seg as representative boundary-aware and transformer methods for transparent objects, but neither is included in the experimental comparison. Since the paper claims that the proposed edge-aware design improves boundary F-score and is suitable for real-time use, a direct comparison with LBSNet (and ideally Trans2Seg, with latency measured) is needed to substantiate the advantage over existing transparent-object-specific architectures.

minor comments (5)

[Abstract] The text contains a spacing error: "constructLabGlass-IS" should be "construct LabGlass-IS", and the code/video URLs are broken across lines and should be formatted as proper hyperlinks.
[IV-E] Please clarify what the reported "average per-frame latency" of 73.70 ms includes, since it is an order of magnitude larger than the 7.1 ms segmentation inference time and does not match the earlier timing claim.
[III-B] In Eq. (4), define |M_i| explicitly as the number of pixels in the mask, and in Eq. (6) clarify that R_i and C_i are expressed in the robot base frame.
[Table I] The abstract's phrase "closest accuracy competitor" should be defined explicitly; if it refers to PointRend in terms of mAP50:95, the text should state that this is the accuracy metric used for the parameter comparison.
[Dataset] Since LabGlass-IS is introduced as a new benchmark, please provide a dataset download link or a clear statement of release plans in the final version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the paper's evaluations are empirical, and the hand-set cube margin is a reproducibility concern rather than a derivation loop.

full rationale

No significant circularity is present. The perception model is trained and evaluated on the newly constructed LabGlass-IS dataset, with baselines compared under the same split and training settings. The edge branch, SimAM module, and attention fusion are architectural additions whose contributions are tested by ablation, not assumed from a self-citation chain. Multi-view centroid estimation is a standard least-squares ray triangulation from predicted mask centroids, and its output is independently validated against manually measured ground truth in Table III; it does not define its own target. The conservative bounding volume in Section III-C is described as 'conservatively selected to cover the spatial extent of the object while accounting for segmentation uncertainty,' and in Section IV-E the authors state that volumes are 'sized to fully enclose the objects with this margin.' This is an empirical calibration choice made after observing centroid errors. It affects the interpretability and reproducibility of the 93.3% success rate, but the success rate is an experimental outcome rather than a quantity derived from the perception equations, so this is not a fitted parameter renamed as a prediction. There are no load-bearing self-citations, no imported uniqueness theorem, and no known result merely relabeled as a contribution. Under the strict standard that circularity requires a specific reduction of a claimed derivation to its own inputs, the paper is self-contained.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on standard projection and triangulation mathematics, plus two domain assumptions about transparent objects and the mask centroid, and one ad hoc safety margin selection. No new physical entities are introduced.

free parameters (2)

Conservative bounding cube size = unspecified
The cube size around each estimated centroid is chosen by hand to enclose the object and cover segmentation uncertainty (Section III-C); the value is not reported, yet it directly determines whether collisions are avoided.
Minimum ray count N_min = unspecified
Algorithm 1 waits until N_min rays are collected before estimating a 3D centroid. The threshold affects accuracy and latency, but its value is not given.

assumptions (3)

domain assumption Boundary contours of transparent vessels are comparatively reliable visual cues despite refraction and specularity.
Motivates the edge branch (Section III-A). If boundaries are not reliable, the method's core premise fails.
domain assumption Mask centroid back-projection approximates the 3D center of each vessel.
Used in Section III-B for multi-view triangulation; object shape asymmetry may bias the centroid.
ad hoc to paper A conservatively chosen axis-aligned cube centered at the estimated centroid is a safe obstacle model.
Section III-C; the margin is hand-selected, not predicted from uncertainty, so safety is partly encoded in this assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of From Transparent Labware Segmentation to Collision Avoidance: A Real-Time Edge-Aware Perception Pipeline." pith.science (2026). https://pith.science/paper/DNGYDF5S

@misc{pith2026260804769,
  author       = {Pith},
  title        = {Pith review of: From Transparent Labware Segmentation to Collision Avoidance: A Real-Time Edge-Aware Perception Pipeline},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DNGYDF5S}},
  note         = {Machine review of arXiv:2608.04769}
}

read the original abstract

This paper presents an edge-aware instance segmentation framework that enables real-time robotic collision avoidance with transparent laboratory glassware using purely visual perception. Transparent vessels defy conventional segmentation due to refraction, specular reflection, and the absence of stable interior texture, yet their boundary contours remain comparatively reliable visual cues. Exploiting this observation, we augment a one-stage real-time instance segmentation backbone with a lightweight edge-detection branch, edge-guided attention fusion, and a parameter-free SimAM module, and further construct LabGlass-IS, a 3485-image, 21-category instance segmentation dataset of real laboratory glassware. The enhanced model achieves the highest Boundary F-score of 97.80 among compared methods, outperforming the YOLO-prompted FastSAM framework by 18.93 BF points. Furthermore, it maintains an inference speed of 7.1ms per frame and requires only 2.85% of the parameters of the closest accuracy competitor. Multi-view triangulation of mask centroids further provides 3D positions for conservative bounding-volume collision constraints. Real-robot trials achieve a 93.3% collision avoidance success rate, indicating the feasibility of the proposed perception-to-action pipeline for robot collision avoidance among fragile transparent objects. Our code is available at https://github.com/havishamy/TransYOLO_3D. Our video is available at https://havishamy.github.io/paper-videos/.

Figures

Figures reproduced from arXiv: 2608.04769 by the authors.

**Figure 1.** Pipeline of perception-driven robotic collision avoidance for transparent labware. The eye-in-hand camera on the robotic arm captures visual data, the 3D centroid of each vessel is derived from the segmentation mask centroid via multi-view triangulation, and a conservative 3D bounding box centered at the centroid is built as the collision avoidance constraint for robotic manipulation. consumption due to dense atten… view at source ↗

**Figure 2.** Overview of the proposed method. It comprises a hierarchical backbone with SimAM attention for multi-scale feature [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** From left to right: RGB images, intermediate feature [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Samples of the 21 laboratory apparatus types from our [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Qualitative comparison on transparent glassware instance segmentation. Zoomed-in regions mainly highlight the [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Real-robot obstacle avoidance experiment. The red [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 24 canonical work pages

[1]

Robotic perception of transparent objects: A review,

J. Jiang, G. Cao, J. Deng, T.-T. Do, and S. Luo, “Robotic perception of transparent objects: A review,”IEEE Trans. Artif. Intell., vol. 5, no. 6, pp. 2547–2567, 2023

work page 2023
[2]

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,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 4649–4658

work page 2021
[3]

Super lidar intensity for robotic perception,

W. Gao, J. Zhang, M. Zhao, Z. Zhang, S. Kong, M. Ghaffari, D. Song, C. Xu, and H. Kong, “Super lidar intensity for robotic perception,” IEEE Robot. Autom. Lett., 2026

work page 2026
[4]

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

J. Lin, Y .-H. Yeung, S. Ye, and R. W. H. Lau, “Leveraging rgb-d data with cross-modal context mining for glass surface detection,” inProc. AAAI Conf. Artif. Intell., vol. 39, no. 5, 2025, pp. 5254–5261

work page 2025
[5]

Lbsnet: Lightweight joint boundary detection and semantic segmentation for transparent and reflective objects,

L. Tong, K. Qian, and X. Jing, “Lbsnet: Lightweight joint boundary detection and semantic segmentation for transparent and reflective objects,”IEEE Robot. Autom. Lett., vol. 10, no. 2, pp. 955–962, 2024

work page 2024
[6]

Segmenting Transparent Object in the Wild with Transformer

E. Xie, W. Wang, W. Wang, P. Sun, H. Xu, D. Liang, and P. Luo, “Segmenting transparent object in the wild with transformer,”arXiv preprint arXiv:2101.08461, 2021

work page Pith review arXiv 2021
[7]

Segment anything,

A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo,et al., “Segment anything,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023, pp. 4015–4026

work page 2023
[8]

Seg- menting transparent objects in the wild,

E. Xie, W. Wang, W. Wang, M. Ding, C. Shen, and P. Luo, “Seg- menting transparent objects in the wild,” inEur. Conf. Comput. Vis. (ECCV). Springer, 2020, pp. 696–711

work page 2020

Show all 27 references

[9]

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,” inIEEE Int. Conf. Robot. Autom. (ICRA). IEEE, 2020, pp. 3634–3642

2020
[10]

A benchmark dataset in chemical apparatus: recognition and detection,

L. Zou, Z.-S. Ding, S.-Y . Ran, Z.-Z. Wu, Y .-S. Wei, Z.-H. He, and X.- F. Wang, “A benchmark dataset in chemical apparatus: recognition and detection,”Multimed. Tools Appl., vol. 83, no. 9, pp. 26 419–26 437, 2024

2024
[11]

Computer vision for recognition of materials and vessels in chemistry lab settings and the vector-labpics data set,

S. Eppel, H. Xu, M. Bismuth, and A. Aspuru-Guzik, “Computer vision for recognition of materials and vessels in chemistry lab settings and the vector-labpics data set,”ACS Cent. Sci., vol. 6, no. 10, pp. 1743– 1752, 2020

2020
[12]

Yolact: Real-time instance segmentation,

D. Bolya, C. Zhou, F. Xiao, and Y . J. Lee, “Yolact: Real-time instance segmentation,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 9157–9166

2019
[13]

You only look once-object detection mod- els: a review,

A. Nazir and M. A. Wani, “You only look once-object detection mod- els: a review,” inInt. Conf. Comput. Sustain. Global Dev. (INDIACom). IEEE, 2023, pp. 1088–1095

2023
[14]

Yolact++: Better real- time instance segmentation,

D. Bolya, C. Zhou, F. Xiao, and Y . J. Lee, “Yolact++: Better real- time instance segmentation,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 2, pp. 1108–1121, 2020

2020
[15]

Solo: Segmenting objects by locations,

X. Wang, T. Kong, C. Shen, Y . Jiang, and L. Li, “Solo: Segmenting objects by locations,” inEur. Conf. Comput. Vis. (ECCV). Springer, 2020, pp. 649–665

2020
[16]

Masked-attention mask transformer for universal image segmen- tation,

B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, “Masked-attention mask transformer for universal image segmen- tation,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 1290–1299

2022
[17]

Fast segment anything,

X. Zhao, W. Ding, Y . An, Y . Du, T. Yu, M. Li, M. Tang, and J. Wang, “Fast segment anything,”arXiv preprint arXiv:2306.12156, 2023

2023 arXiv
[18]

Collision avoidance for elliptical agents with control barrier function utilizing supporting lines,

K. Nishimoto, R. Funada, T. Ibuki, and M. Sampei, “Collision avoidance for elliptical agents with control barrier function utilizing supporting lines,” inAmer. Control Conf. (ACC). IEEE, 2022, pp. 5147–5153

2022
[19]

Predictive collision avoidance for the dynamic window approach,

M. Missura and M. Bennewitz, “Predictive collision avoidance for the dynamic window approach,” inIEEE Int. Conf. Robot. Autom. (ICRA). IEEE, 2019, pp. 8620–8626

2019
[20]

Multi-joint active collision avoidance for robot based on depth visual perception,

H. Li, X. Wang, X. Huang, Y . Ma, and Z. Jiang, “Multi-joint active collision avoidance for robot based on depth visual perception,” IEEE/CAA J. Autom. Sinica, vol. 9, no. 12, pp. 2186–2189, 2022

2022
[21]

Mvtrans: Multi-view perception of transparent objects,

Y . R. Wang, Y . Zhao, H. Xu, S. Eppel, A. Aspuru-Guzik, F. Shkurti, and A. Garg, “Mvtrans: Multi-view perception of transparent objects,” arXiv preprint arXiv:2302.11683, 2023

2023 arXiv
[22]

Rftrans: Leveraging refractive flow of transparent objects for surface normal estimation and manipulation,

T. Tang, J. Liu, J. Zhang, H. Fu, W. Xu, and C. Lu, “Rftrans: Leveraging refractive flow of transparent objects for surface normal estimation and manipulation,”IEEE Robot. Autom. Lett., vol. 9, no. 4, pp. 3735–3742, 2024

2024
[23]

Cagt: Sim-to-real depth completion with interactive embedding aggregation and geometry awareness for transparent objects,

X. Jing, K. Qian, and M. Vincze, “Cagt: Sim-to-real depth completion with interactive embedding aggregation and geometry awareness for transparent objects,”IEEE Trans. Circuits Syst. Video Technol., vol. 35, no. 7, pp. 6656–6670, 2025

2025
[24]

Torm: Transparent objects reconstruction and manipulation with multi-view segmentation,

Q. Qiao, F. Lin, H. Zhao, B. Xu, Z. Chen, D. Xu, and P. Lu, “Torm: Transparent objects reconstruction and manipulation with multi-view segmentation,”IEEE Robot. Autom. Lett., vol. 11, no. 1, pp. 242–249, 2026

2026
[25]

Heapgrasp: Hand-eye active perception to grasp objects with diverse optical properties,

G. Kennis and S. Arai, “Heapgrasp: Hand-eye active perception to grasp objects with diverse optical properties,”IEEE Robot. Autom. Lett., vol. 11, no. 3, pp. 3206–3213, 2026

2026
[26]

Simam: A simple, parameter-free attention module for convolutional neural networks,

L. Yang, R.-Y . Zhang, L. Li, and X. Xie, “Simam: A simple, parameter-free attention module for convolutional neural networks,” in Int. Conf. Mach. Learn. (ICML). PMLR, 2021, pp. 11 863–11 874

2021
[27]

Pointrend: Image segmentation as rendering,

A. Kirillov, Y . Wu, K. He, and R. Girshick, “Pointrend: Image segmentation as rendering,” inIEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 9796–9805

2020

Pith tools

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

[1] [1]

Robotic perception of transparent objects: A review,

J. Jiang, G. Cao, J. Deng, T.-T. Do, and S. Luo, “Robotic perception of transparent objects: A review,”IEEE Trans. Artif. Intell., vol. 5, no. 6, pp. 2547–2567, 2023

work page 2023

[2] [2]

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,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 4649–4658

work page 2021

[3] [3]

Super lidar intensity for robotic perception,

W. Gao, J. Zhang, M. Zhao, Z. Zhang, S. Kong, M. Ghaffari, D. Song, C. Xu, and H. Kong, “Super lidar intensity for robotic perception,” IEEE Robot. Autom. Lett., 2026

work page 2026

[4] [4]

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

J. Lin, Y .-H. Yeung, S. Ye, and R. W. H. Lau, “Leveraging rgb-d data with cross-modal context mining for glass surface detection,” inProc. AAAI Conf. Artif. Intell., vol. 39, no. 5, 2025, pp. 5254–5261

work page 2025

[5] [5]

Lbsnet: Lightweight joint boundary detection and semantic segmentation for transparent and reflective objects,

L. Tong, K. Qian, and X. Jing, “Lbsnet: Lightweight joint boundary detection and semantic segmentation for transparent and reflective objects,”IEEE Robot. Autom. Lett., vol. 10, no. 2, pp. 955–962, 2024

work page 2024

[6] [6]

Segmenting Transparent Object in the Wild with Transformer

E. Xie, W. Wang, W. Wang, P. Sun, H. Xu, D. Liang, and P. Luo, “Segmenting transparent object in the wild with transformer,”arXiv preprint arXiv:2101.08461, 2021

work page Pith review arXiv 2021

[7] [7]

Segment anything,

A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo,et al., “Segment anything,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023, pp. 4015–4026

work page 2023

[8] [8]

Seg- menting transparent objects in the wild,

E. Xie, W. Wang, W. Wang, M. Ding, C. Shen, and P. Luo, “Seg- menting transparent objects in the wild,” inEur. Conf. Comput. Vis. (ECCV). Springer, 2020, pp. 696–711

work page 2020