ArrowPose: Segmentation, Detection, and 5 DoF Pose Estimation Network for Colorless Point Clouds

Frederik Hagelskjaer

REVIEW 4 major objections 5 minor 29 references

ArrowPose: Segmentation, Detection, and 5 DoF Pose Estimation Network for Colorless Point Clouds

T0 review · 4 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 A depth-only network that detects and estimates 5 DoF poses of cylindrically symmetric objects by regressing center and top points reports state-of-the-art results on the IC-BIN benchmark.

desk verdict A fast, clever depth-only 5-DoF pose network with a genuinely new architecture, but the state-of-the-art claim rests on a thin official-score margin from a tiny two-object benchmark. read the letter →

arxiv 2506.08699 v1 pith:4G7HURA6 submitted 2025-06-10 cs.CV

Frederik Hagelskjaer This is my paper

classification cs.CV

keywords 5-DoFposeestimationdepth-onlyobjectdetectionpointcloudcylindricalsymmetryinstancesegmentationIC-BINbenchmarksynthetic-to-realtransferbinpicking

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 tries to establish that a single network operating on depth-only point clouds can perform detection, instance segmentation, and 5 degree-of-freedom pose estimation for rotationally symmetric industrial objects, without any color information. The network predicts, for each point on an object, an offset to the object's center and to a point on its symmetry axis; clustering these predictions produces object detections, and the center-to-top vector yields the pose. The method is trained exclusively on synthetic depth images, with random holes, noise, blur, cropping, and edge removal used to bridge the gap to real sensors. On the real IC-BIN benchmark it reports an average score of 71.8 versus 57.6 for the previous best method and an AR of 0.417 versus 0.396 for the best depth-only PPF variant. If these results hold, robots could localize parts such as cups, bearings, and screws in about 250 ms per scene using only a depth sensor.

What carries the argument

The load-bearing machinery is the pair of center and top offset regressions, combined with a non-maximum suppression clustering step. Each object point votes for the object's center and for a top point on the symmetry axis; NMS with a threshold of half the shortest bounding-box side groups votes that belong to the same instance, and the surviving center plus the averaged center-to-top vector gives the 5 DoF pose. This is supported by two EdgeConv variants: LocalEdgeConv, which uses only neighbor differences to remove global position noise, and SpatialEdgeConv, which reintroduces spatial neighbor differences in later layers. To handle 65,536 points, neighbor indices are precomputed on the CPU and the cloud is randomly down-sampled across four layers to 2,048 points. A final Segmentation-ICP step refines the pose by fitting only the segmented scene points to the object model, which the paper reports as both faster and more accurate than classic ICP.

What would settle it

Run the published training code on the same BOP synthetic IC-BIN data and test on the real IC-BIN set using the BOP AR_VSD protocol; if the reproduced average AR_VSD falls at or below the best PPF baseline's 0.510, the claim that the method outperforms all depth-only methods is overturned.

Watch

Extended reading notes

Core claim

The central claim is that a 5 DoF pose can be recovered from a colorless point cloud by regressing two per-point offsets: one to the object's bounding-box center and one to the top of its symmetry axis, then reading the pose directly from the center point and the center-to-top vector. The paper argues this is enough for the large class of cylindrically symmetric industrial objects, and that for such objects attempting to estimate the missing sixth degree of freedom is estimating aleatoric noise rather than useful signal. The network is built from EdgeConv layers with local and spatial feature differences and a PointNet++-style pyramid, which lets it consume 65,536 points and still segment and detect in one pass. The paper's experiments on IC-BIN are presented as evidence that this depth-only approach beats all colorless baselines, particularly on the rotationally symmetric Coffee object where the 5 DoF formulation gives the largest gain.

Load-bearing premise

The entire accuracy story depends on the trained network transferring from perfect synthetic depth images to real depth cameras through random augmentations; if a real sensor's noise or missing-edge pattern differs from that augmentation mix, the reported 71.8 average score is not guaranteed.

Editorial extensions

If this is right

On the IC-BIN benchmark, the method raises the average pose score from 57.6 (previous best) to 71.8, with the rotationally symmetric Coffee object gaining almost 20 points.
Under the BOP metric, the method reaches AR 0.417 and AR_VSD 0.561, beating the best depth-only PPF variant at AR 0.396 and AR_VSD 0.510, and the VSD gain is larger than the AR gain.
A full scene with ten objects is processed in under 0.250 s for 5 DoF poses, so the pipeline can run at interactive rates on a laptop GPU; the 6 DoF extension costs 2.68 s.
Using the predicted instance segmentation to constrain ICP makes refinement faster (0.037 s vs 0.365 s for classic ICP) and raises the Coffee score from 79.6 to 83.6.
Increasing the k-NN receptive field from 20 to 40 neighbors improves the overall score from 71.8 to 73.9, mainly on the larger Juice object, at the cost of slower neighbor computation.

Reading between the lines

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

The paper leaves implicit that the same center-plus-top scheme applies to any object with a defined symmetry axis, so idlers, bearings, and screws could be handled with new synthetic data and a new clustering threshold.
A testable extension not in the paper: train on synthetic depth from one sensor model and evaluate on a physically different depth camera with no re-training; the drop in AR_VSD would quantify how much of the reported accuracy rests on the specific augmentation mix.
Since the 6 DoF extension is a post-hoc 10-degree search rather than a learned prediction, the paper's 2.68 s 6 DoF runtime is likely dominated by that search; predicting a tangent direction in the same forward pass could in principle restore the 250 ms speed for non-symmetric objects.

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

Frederik,

I read ArrowPose and the stress-test note. The stress-test is right: the headline 'state-of-the-art' claim is shakier than the abstract suggests. On the BOP AR, the gain over the best PPF baseline is 0.417 vs 0.396, and the paper's own Table II shows the real advantage is in AR_VSD (0.561 vs 0.510). Since AR averages VSD, MSSD, and MSPD, the unreported sub-metrics are essentially tied. That doesn't sink the paper, but it means the 'outperforms all colorless methods' claim is only true for one metric, not the official average.

What's genuinely good: the architecture is a clean, practical combination of DGCNN's EdgeConv with PointNet++'s pyramid, plus the two small adaptations (LocalEdgeConv, SpatialEdgeConv) that make sense for object detection in large point clouds. Precomputing neighbors on the CPU is a simple idea that pays off in runtime. The center-and-top offset regression is a neat way to get 5-DoF pose without predicting a full rotation, and the Segmentation-ICP refinement is a legitimate improvement over classic ICP, both faster and more accurate. The paper is clearly written, the experiments are reproducible in principle (code is promised), and the ablations on neighbor count and ICP give useful information.

The soft spots are real but proportionate. The evaluation is one benchmark, two objects, roughly 180 test images, no error bars, no per-scene breakdown. The transfer from synthetic depth to real sensor data is handled with augmentations, and the results suggest it works, but a single small benchmark can't tell you how robust that is. The 250 ms runtime is on a laptop GPU, which is fine, but the comparison methods' runtimes aren't reported. I'd also like to see the per-object numbers for all three BOP sub-metrics, not just AR and AR_VSD.

Who this is for: anyone working on depth-only pose estimation for bin picking or industrial manipulation, especially for rotationally symmetric objects. It's a useful method paper, not a breakthrough claim. It deserves a serious referee: the engineering is solid, the method is novel enough, and the claims are testable. My recommendation is to send it to review, but the authors should be asked to report the full BOP sub-metrics, add error bars or multiple runs, and tone down 'state-of-the-art' to 'competitive with state-of-the-art on IC-BIN.'

Best, [Your name]

Referee Report

4 major / 5 minor

Summary. The paper proposes ArrowPose, a network for segmentation, detection, and 5 DoF pose estimation of cylindrically symmetric objects from colorless point clouds. The network predicts object centers and top points from a hierarchical EdgeConv architecture, clusters center predictions to obtain instance segmentations, and derives the pose by a geometric construction from the center-top vector. A segmentation-ICP refinement and an optional 6 DoF extension by z-axis search are described. The network is trained on synthetic BOP data with augmentation and evaluated on the IC-BIN benchmark, where the authors report state-of-the-art performance among depth-only methods and a runtime of about 0.25 s.

Significance. If the reported results are confirmed, the method is a useful contribution to industrial bin picking: it processes large point clouds without color information, is fast, and is clearly described with open code. The ablations on the number of neighbors and on Segmentation-ICP versus classic ICP are informative, and the runtime breakdown is a strength. However, the central state-of-the-art claim currently rests on a small benchmark with two objects, a thin margin in the official BOP AR score, and incomplete reporting of the BOP sub-metrics, so the empirical support does not yet match the strength of the claim.

major comments (4)

[III-A1 (Table II)] The state-of-the-art claim rests on an overall AR gain of only 0.021 over the best PPF baseline (0.417 versus 0.396), while the larger gain is concentrated in AR_VSD (0.561 versus 0.510). The paper does not report AR_MSSD or AR_MSPD, so the reader cannot verify whether the official BOP average is genuinely improved or whether the advantage is confined to the symmetry-invariant VSD term. Please report all three BOP sub-metrics, and ideally per-object and per-scene scores.
[III-A] IC-BIN contains only two objects and roughly 180 test images, and all results come from a single run without error bars. A 0.021 difference in AR and a 0.051 difference in AR_VSD are statistically fragile in this setting. Please provide variance estimates, such as multiple training runs, per-scene breakdowns, or per-object breakdowns, and discuss the stability of the reported margins.
[II-B and III-A] The synthetic-to-real transfer is load-bearing: the network is trained on perfect synthetic depth projections and generalizes only through augmentations (random holes, noise, blur, cropping, edge removal). The paper provides no failure-case analysis or per-scene variation study, so it does not demonstrate that the augmentation strategy covers real sensor noise. Please add qualitative failure cases and a discussion of which augmentations are necessary, or test on a second depth sensor.
[III-A1 (Table II)] The comparison set is limited to Drost-CVPR10-3D-Only, Vidal-Sensors18, ZTE PPF, and SMC [5]. If no recent learning-based depth-only detectors are available on IC-BIN, the abstract and conclusion should scope the claim as 'best among the compared methods' rather than 'outperforms all colorless methods' or 'state-of-the-art'.

minor comments (5)

[Abstract and Conclusion] The abstract and conclusion state 'outperforms all colorless methods' and 'state-of-the-art performance,' but the experimental comparison covers only four methods; please align the wording with the actual comparison set.
[II-A5 (Eq. 4)] Equation (4) defines the orientation vector as a norm, |p_center - p_top|, which is a scalar, while Equations (5)-(7) treat it as a vector. The notation should be corrected, for example by defining a unit vector rather than a norm.
[II-A1] The sentence 'However, as the global position information is recorded the relative position between points is unknown in the is kept' is grammatically garbled and should be rewritten.
[II-A1 and Fig. 2] In the text, the PointNet++ reference [18] should likely be [14], and the Fig. 2 caption says 'for each n class' where 'n classes' is intended.
[III-B] The runtime table is measured on a laptop GPU; please state whether these timings are representative of the target deployment setting and clarify whether they include the 6 DoF search or only the 5 DoF pipeline.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: pose is derived by geometry from network outputs trained on synthetic data and evaluated on an external benchmark.

full rationale

The claimed derivation chain is: the network predicts per-point center and top offsets; NMS clustering yields object centers; the 5 DoF pose is computed by subtracting top from center and taking cross products (Eq. 4-7); Segmentation-ICP refines using the segmentation; the 6 DoF extension searches z-rotations scored by a depth-overlap projection. None of these steps use IC-BIN test labels or fit parameters to the evaluation set. The network is trained on synthetic BOP depth images with augmentations; the IC-BIN evaluation is an external real-depth benchmark. The self-citations [16], [21], [28] are implementation or baseline references: [28] is a previously reported score used only for comparison, [21] supplies a projection scoring subroutine, and [16] supports a feasibility claim about adjacency matrices. None are used as a uniqueness theorem or to force the outcome. The state-of-the-art claim rests on measured AR and AR_VSD values computed by the BOP protocol. Therefore no circular reduction exists.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method relies on standard deep learning with hand-chosen hyperparameters and domain assumptions about object symmetry and sensor-noise transfer. No new physical entities are introduced.

free parameters (5)

NMS threshold distance (half shortest bounding-box side)
Used to cluster center predictions into instances; hand-chosen from the object geometry in Section II-A4.
Number of k-NN neighbors = 20
Ablation shows 10 degrades performance (52.5 vs 71.8) and 40 improves marginally to 73.9 at higher runtime; 20 is the chosen trade-off in Section III-D.
Point cloud size and downsampling schedule = 65,536 to 2,048
Chosen for fast processing on CPU in Section II-A3; the paper asserts dense sampling retains information but does not ablate this choice.
ICP overlap threshold = 80%
Used in 6 DoF refinement to discard poor rotations in Section II-A7; value chosen by hand.
Z-axis rotation step = 10 degrees
Search resolution for 6 DoF in Section II-A7; finer steps would increase the 2.68 s runtime.

assumptions (4)

domain assumption The objects of interest have cylindrical symmetry for 5 DoF pose estimation.
The method estimates orientation only along the symmetry axis; stated in Section I.
domain assumption Synthetic depth images from BOP, after the described augmentations, are representative of real sensor depth.
Training uses synthetic depth; generalization to real depth rests on this assumption, evaluated only on IC-BIN.
domain assumption Random downsampling preserves enough information for detection and segmentation.
The paper argues dense sampling plus hierarchical downsampling 'almost guarantees' information retention (Section II-A3); this is not formally verified.
domain assumption Neighbor computation on the CPU based on spatial distance is sufficient, omitting feature-space neighbors.
Adaptation from DGCNN to scale to 65,536 points; may reduce receptive field but is essential for the reported runtime.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ArrowPose: Segmentation, Detection, and 5 DoF Pose Estimation Network for Colorless Point Clouds." pith.science (2026). https://pith.science/paper/4G7HURA6

@misc{pith2026250608699,
  author       = {Pith},
  title        = {Pith review of: ArrowPose: Segmentation, Detection, and 5 DoF Pose Estimation Network for Colorless Point Clouds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4G7HURA6}},
  note         = {Machine review of arXiv:2506.08699}
}

read the original abstract

This paper presents a fast detection and 5 DoF (Degrees of Freedom) pose estimation network for colorless point clouds. The pose estimation is calculated from center and top points of the object, predicted by the neural network. The network is trained on synthetic data, and tested on a benchmark dataset, where it demonstrates state-of-the-art performance and outperforms all colorless methods. The network is able to run inference in only 250 milliseconds making it usable in many scenarios. Project page with code at arrowpose.github.io

Figures

Figures reproduced from arXiv: 2506.08699 by the authors.

**Figure 2.** Network Architecture: The network input is a point cloud and precomputed neighbor indices. For each layer neighbor differences are computed and processed, while the point cloud is downscaled. Finally, the features for the last two layers are concatenated, ⊕, and processed by an MLP. Using this, the segmentation prediction is computed for each n class. The MLP output is concatenated with the earlier concatenated fea… view at source ↗

**Figure 3.** Left (a): visualization of the density of the sampling of the networks [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Visualization of the object detection and segmentation process. (a) [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

**Figure 5.** Figure 5: Examples of pose estimation results on the test dataset. To the left [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 22 canonical work pages

[5]

Bayesian inference for cad-based pose estimation on depth images for robotic manipulation,

T. Redick, “Bayesian inference for cad-based pose estimation on depth images for robotic manipulation,” Ph.D. dissertation, Dissertation, Rheinisch-Westf¨alische Technische Hochschule Aachen, 2024, 2024

work page 2024
[1]

Bop: Benchmark for 6d object pose estimation,

T. Hodan, F. Michel, E. Brachmann, W. Kehl, A. GlentBuch, D. Kraft, B. Drost, J. Vidal, S. Ihrke, X. Zabulis et al. , “Bop: Benchmark for 6d object pose estimation,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 19–34

work page 2018
[2]

BOP Challenge 2022 on Detection, Segmentation and Pose Estimation of Specific Rigid Objects

M. Sundermeyer, T. Hodan, Y . Labbe, G. Wang, E. Brachmann, B. Drost, C. Rother, and J. Matas, “Bop challenge 2022 on detection, segmen- tation and pose estimation of specific rigid objects,” arXiv preprint arXiv:2302.13075, 2023

work page Pith review arXiv 2022
[3]

Bop challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects,

T. Hodan, M. Sundermeyer, Y . Labbe, V . N. Nguyen, G. Wang, E. Brach- mann, B. Drost, V . Lepetit, C. Rother, and J. Matas, “Bop challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 5610–5619

work page 2023
[4]

Model globally, match locally: Efficient and robust 3d object recognition,

B. Drost, M. Ulrich, N. Navab, and S. Ilic, “Model globally, match locally: Efficient and robust 3d object recognition,” in 2010 IEEE computer society conference on computer vision and pattern recognition . Ieee, 2010, pp. 998–1005

work page 2010
[6]

A hybrid approach for 6dof pose estimation,

R. K ¨onig and B. Drost, “A hybrid approach for 6dof pose estimation,” in European Conference on Computer Vision . Springer, 2020, pp. 700– 706

work page 2020
[7]

Pointpillars: Fast encoders for object detection from point clouds,

A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 12 697–12 705

2019
[8]

Autonomous driving system: A comprehensive survey,

J. Zhao, W. Zhao, B. Deng, Z. Wang, F. Zhang, W. Zheng, W. Cao, J. Nan, Y . Lian, and A. F. Burke, “Autonomous driving system: A comprehensive survey,” Expert Systems with Applications , vol. 242, p. 122836, 2024

work page 2024

Show all 29 references

[9]

V oxelnext: Fully sparse voxelnet for 3d object detection and tracking,

Y . Chen, J. Liu, X. Zhang, X. Qi, and J. Jia, “V oxelnext: Fully sparse voxelnet for 3d object detection and tracking,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 21 674–21 683

2023
[10]

Assembly challenge: a robot competition of the industrial robotics category, world robot summit–summary of the pre-competition in 2018,

Y . Yokokohji, Y . Kawai, M. Shibata, Y . Aiyama, S. Kotosaka, W. Ue- mura, A. Noda, H. Dobashi, T. Sakaguchi, and K. Yokoi, “Assembly challenge: a robot competition of the industrial robotics category, world robot summit–summary of the pre-competition in 2018,” Advanced Robot...

2018
[11]

Recovering 6d object pose and predicting next-best-view in the crowd,

A. Doumanoglou, R. Kouskouridas, S. Malassiotis, and T.-K. Kim, “Recovering 6d object pose and predicting next-best-view in the crowd,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3583–3592

2016
[12]

A workpiece grasp detection method based on 3d object detection,

H. Li, L. Duan, Q. Wang, Y . Zhang, and B. Ye, “A workpiece grasp detection method based on 3d object detection,” Industrial Robot: the international journal of robotics research and application , 2025

2025
[13]

Instance segmentation based 6d pose estimation of industrial objects using point clouds for robotic bin- picking,

C. Zhuang, S. Li, and H. Ding, “Instance segmentation based 6d pose estimation of industrial objects using point clouds for robotic bin- picking,” Robotics and Computer-Integrated Manufacturing , vol. 82, p. 102541, 2023

2023
[14]

Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,” Advances in neural information processing systems , vol. 30, 2017

2017
[15]

Dynamic graph cnn for learning on point clouds,

Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon, “Dynamic graph cnn for learning on point clouds,” ACM Transactions on Graphics , 2019

2019
[16]

Deep learning classification of large-scale point clouds: A case study on cuneiform tablets,

F. Hagelskjær, “Deep learning classification of large-scale point clouds: A case study on cuneiform tablets,” in 2022 IEEE International Con- ference on Image Processing (ICIP) . IEEE, 2022, pp. 826–830

2022
[17]

Nearest neighbor pattern classification,

T. Cover and P. Hart, “Nearest neighbor pattern classification,” IEEE transactions on information theory , vol. 13, no. 1, pp. 21–27, 1967

1967
[18]

Pointnet: Deep learning on point sets for 3d classification and segmentation,

C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 652–660

2017
[19]

An adjustable farthest point sampling method for approximately-sorted point cloud data,

J. Li, J. Zhou, Y . Xiong, X. Chen, and C. Chakrabarti, “An adjustable farthest point sampling method for approximately-sorted point cloud data,” in 2022 IEEE Workshop on Signal Processing Systems (SiPS) , 2022, pp. 1–6

2022
[20]

Least-squares fitting of two 3-d point sets,

K. S. Arun, T. S. Huang, and S. D. Blostein, “Least-squares fitting of two 3-d point sets,” IEEE Transactions on pattern analysis and machine intelligence, no. 5, pp. 698–700, 1987

1987
[21]

Bridging the reality gap for pose estimation networks using sensor-based domain randomization,

F. Hagelskjær and A. G. Buch, “Bridging the reality gap for pose estimation networks using sensor-based domain randomization,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 935–944

2021
[22]

Bop challenge 2020 on 6d object localization,

T. Hoda ˇn, M. Sundermeyer, B. Drost, Y . Labb ´e, E. Brachmann, F. Michel, C. Rother, and J. Matas, “Bop challenge 2020 on 6d object localization,” in Computer Vision–ECCV 2020 Workshops: Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16 . Springer, 2020, pp. 577–594

2020
[23]

Photorealistic image synthesis for object instance detection,

T. Hoda ˇn, V . Vineet, R. Gal, E. Shalev, J. Hanzelka, T. Connell, P. Urbina, S. N. Sinha, and B. Guenter, “Photorealistic image synthesis for object instance detection,” in 2019 IEEE international conference on image processing (ICIP) . IEEE, 2019, pp. 66–70

2019
[24]

A computational approach to edge detection,

J. Canny, “A computational approach to edge detection,” IEEE Transac- tions on pattern analysis and machine intelligence , no. 6, pp. 679–698, 1986

1986
[25]

Learning to model the tail,

Y .-X. Wang, D. Ramanan, and M. Hebert, “Learning to model the tail,” Advances in neural information processing systems , vol. 30, 2017

2017
[26]

Latent-class hough forests for 3d object detection and pose estimation,

A. Tejani, D. Tang, R. Kouskouridas, and T.-K. Kim, “Latent-class hough forests for 3d object detection and pose estimation,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13 . Springer, 2014, pp. 462–477

2014
[27]

Rotational subgroup voting and pose clustering for robust 3d object recognition,

A. G. Buch, L. Kiforenko, and D. Kraft, “Rotational subgroup voting and pose clustering for robust 3d object recognition,” in 2017 IEEE International Conference on Computer Vision (ICCV) . IEEE, 2017, pp. 4137–4145

2017
[28]

Bayesian optimization of 3d feature parameters for 6d pose estimation,

F. Hagelskjær, N. Kr ¨uger, and A. G. Buch, “Bayesian optimization of 3d feature parameters for 6d pose estimation,” in 14th International Con- ference on Computer Vision Theory and Applications . SCITEPRESS Digital Library, 2019, pp. 135–142

2019
[29]

Implicit-pdf: Non-parametric representation of probability distributions on the rotation manifold,

K. Murphy, C. Esteves, V . Jampani, S. Ramalingam, and A. Makadia, “Implicit-pdf: Non-parametric representation of probability distributions on the rotation manifold,” arXiv preprint arXiv:2106.05965 , 2021

2021 arXiv

Pith tools

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

[1] [5]

Bayesian inference for cad-based pose estimation on depth images for robotic manipulation,

T. Redick, “Bayesian inference for cad-based pose estimation on depth images for robotic manipulation,” Ph.D. dissertation, Dissertation, Rheinisch-Westf¨alische Technische Hochschule Aachen, 2024, 2024

work page 2024

[2] [1]

Bop: Benchmark for 6d object pose estimation,

T. Hodan, F. Michel, E. Brachmann, W. Kehl, A. GlentBuch, D. Kraft, B. Drost, J. Vidal, S. Ihrke, X. Zabulis et al. , “Bop: Benchmark for 6d object pose estimation,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 19–34

work page 2018

[3] [2]

BOP Challenge 2022 on Detection, Segmentation and Pose Estimation of Specific Rigid Objects

M. Sundermeyer, T. Hodan, Y . Labbe, G. Wang, E. Brachmann, B. Drost, C. Rother, and J. Matas, “Bop challenge 2022 on detection, segmen- tation and pose estimation of specific rigid objects,” arXiv preprint arXiv:2302.13075, 2023

work page Pith review arXiv 2022

[4] [3]

Bop challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects,

T. Hodan, M. Sundermeyer, Y . Labbe, V . N. Nguyen, G. Wang, E. Brach- mann, B. Drost, V . Lepetit, C. Rother, and J. Matas, “Bop challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 5610–5619

work page 2023

[5] [4]

Model globally, match locally: Efficient and robust 3d object recognition,

B. Drost, M. Ulrich, N. Navab, and S. Ilic, “Model globally, match locally: Efficient and robust 3d object recognition,” in 2010 IEEE computer society conference on computer vision and pattern recognition . Ieee, 2010, pp. 998–1005

work page 2010

[6] [6]

A hybrid approach for 6dof pose estimation,

R. K ¨onig and B. Drost, “A hybrid approach for 6dof pose estimation,” in European Conference on Computer Vision . Springer, 2020, pp. 700– 706

work page 2020

[7] [7]

Pointpillars: Fast encoders for object detection from point clouds,

A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 12 697–12 705

2019

[8] [8]

Autonomous driving system: A comprehensive survey,

J. Zhao, W. Zhao, B. Deng, Z. Wang, F. Zhang, W. Zheng, W. Cao, J. Nan, Y . Lian, and A. F. Burke, “Autonomous driving system: A comprehensive survey,” Expert Systems with Applications , vol. 242, p. 122836, 2024

work page 2024