Key-Grid: Unsupervised 3D Keypoints Detection using Grid Heatmap Features

Detecting 3D keypoints with semantic consistency is widely used in many scenarios such as pose estimation, shape registration and robotics. Currently, most unsupervised 3D keypoint detection methods focus on the rigid-body objects. However, when faced with deformable objects, the keypoints they identify do not preserve semantic consistency well. In this paper, we introduce an innovative unsupervised keypoint detector Key-Grid for both the rigid-body and deformable objects, which is an autoencoder framework. The encoder predicts keypoints and the decoder utilizes the generated keypoints to reconstruct the objects. Unlike previous work, we leverage the identified keypoint in formation to form a 3D grid feature heatmap called grid heatmap, which is used in the decoder section. Grid heatmap is a novel concept that represents the latent variables for grid points sampled uniformly in the 3D cubic space, where these variables are the shortest distance between the grid points and the skeleton connected by keypoint pairs. Meanwhile, we incorporate the information from each layer of the encoder into the decoder section. We conduct an extensive evaluation of Key-Grid on a list of benchmark datasets. Key-Grid achieves the state-of-the-art performance on the semantic consistency and position accuracy of keypoints. Moreover, we demonstrate the robustness of Key-Grid to noise and downsampling. In addition, we achieve SE-(3) invariance of keypoints though generalizing Key-Grid to a SE(3)-invariant backbone.

Paper

Similar papers

Peer review

Reviewer LZF74/10 · confidence 2/52024-06-22

Summary

This paper presents a novel unsupervised network for keypoint detection. The method can be applied to handle both rigid and deformable objects. It follows an autoencoder framework where the encoder predicts keypoints and the decoder utilizes the generated keypoints to reconstruct the objects. The main contribution lies in the grid based representation which is suitable to capture the shifted geometric structure of the keypoints brought by deformations.

Strengths

1 This paper presents an unsupervised method for consistent keypoint detection. "Unsupervised" means the method can be applied in larger-scale datasets without 3D annotations. 2 The paper is easy to follow. The writing is good and clear. 3 Extensive experiments show that this method achieves sota performance on many public datasets

Weaknesses

1 The visualized results presented in this paper all feature uniform and complete point clouds. Can this method handle partial point clouds, such as those obtained from back-projecting depth maps? Can it deal with point clouds that contain severe outliers, not just those with some Gaussian noise? This is very important for practical use. 2 In 289, However, if using a SPRIN module [41] which is a SOTA SE(3)-invariant backbone 289 to replace PointNet++ [22] directly, the training process of Key-Grid does not converge. Why? How about using other backbones? like Vector Neurons? Deng, C., Litany, O., Duan, Y., Poulenard, A., Tagliasacchi, A., & Guibas, L. J. (2021). Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 12200-12209). 3 This proposed representation is based on Euclidean distance between the grid point towards each line segment, right? Why not first establish a graph and then aggregate features along the graph? I saw many papers using graph to model deformable objects.

Questions

I have listed some questions in the weakness. The first question is the most important. In L 14, 'Meanwhile, we incorporate the information from each layer of the encoder into the decoder section', Is this your contribution? I recommend to move Fig 6 in the appendix to the main text for clarity.

Rating

4

Confidence

2

Soundness

2

Presentation

3

Contribution

2

Limitations

The author has discussed potential limitations in the paper

Authorsrebuttal2024-08-12

Thank you for your feedback on our paper, particularly regarding the applicability of the Key-Grid to partial and noisy point clouds. We will include this section in the next version of the paper. We greatly appreciate your consideration of the paper as ready for "weak accept". We notice that the rating is not updated. We kindly hope that you update the rating of our paper to reflect the improvements made to the paper.

Reviewer p6XE6/10 · confidence 4/52024-07-09

Summary

This paper presents Key-Grid, an unsupervised keypoint detection network designed for 3D point clouds. Unlike previous methods that emphasize leveraging various priors on 3D structures, this paper converts keypoints into a grid heatmap. This heatmap forms a continuous feature landscape across the entire 3D space, providing richer and more stable geometric descriptions of objects, particularly for deformable objects. Meanwhile, this paper achieves state-of-the-art (SOTA) results on the ShapeNetCoreV2 and ClothesNet datasets.

Strengths

1. The proposed grid heatmap is effective in the task of 3D keypoint detection and represents a novel approach to introducing 3D priors for deformable objects. 2. The authors conducted extensive experiments, achieving SOTA performance on both rigid and deformable objects. 3. The overall design follows the mainstream architecture of 3D keypoint detection, providing convenience for subsequent research in this field.

Weaknesses

Q1. My main concern is whether it can be empirically or theoretically justified why the grid heatmap can represent a dense extension of the skeleton, and why the current method of generating the grid heatmap is reasonable. Could using the density of the 'skeleton' or the density of the point cloud directly represent the grid heatmap instead? Q2. In the ablation study (Table 4), the accuracy drop after "No Grid Heatmap" is minimal, yet the Grid Heatmap is the most significant contribution of this paper. Could the authors explain this phenomenon and provide ablation studies across more categories? Q3. Can this paper adaptively select the number of detected keypoints for different categories of objects? Q4. The quality of figures 1 and 2 could be enhanced. For example, the information content of figure 1 is too minimal to serve as a teaser.

Questions

If the authors address the issues mentioned in the weaknesses, I am willing to increase the rating.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

Yes, the authors discussed the limitations and potential negative societal impact.

Reviewer U8aQ7/10 · confidence 3/52024-07-10

Summary

The authors propose a novel unsupervised method to detect 3D point clouds key points by producing an intermediary heatmap based on a grid and points distances from the skeleton and connected key points. They achieve state-of-the-art accuracy and semantic consistency and easily achieve Se-(3) invariance with minimal adaptation of the method.

Strengths

The method's description, results, and a very in-depth ablation study, together with many supplementary materials, make this paper pleasant to read and easy to understand, with impressive results. Although simple at first glance, the method has the merit of being solid and achieving good results. Great results section and analysis allow the reader to understand in depth the method and results.

Weaknesses

A small neat-picking addition would be to add one or two visual examples where the method does not work super well, to contrast with the excellent results presented in the paper

Questions

Perhaps I missed this point, but how does the author ensure that the key points are semantically the same across different folding/deformation of the objects? ( and give them the same colour)

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

Perhaps the only limitation I can pick up is the consistency across more than two deformations of an object. Would the method perform well over long-term deformations? It ties with showing one or two failure cases of the method.

Reviewer YtUf6/10 · confidence 4/52024-07-13

Summary

Novel PointNet-based autoencoder method called KeyGrid, that predicts semantic keypoints on objects, even when objects are subject to deformations. Similar to previous approaches, keypoints are produced as a linear combination of inputs points, according to a learned weight matrix. The key novelty over related works using PointNet++-based auto-encoders such as SkeletonMerger (SM) [26] is to pass a densely sampled grid of features (3d “point heatmap”) at each decoder layer (which are hierarchical). In this heatmap grid, each cell holds the maximum of the weighted distances to the 'keypoint skeleton segments'. Also, KeyGrid appears to pass richer encoder information to the decoder than the SM work, although the differences here are less clear. The method is shown to outperform previous baselines such as SK3D and SM on ShapeNetV2 dataset (more right objects) and on ClothesNet (deformable objects).

Strengths

- Beats KeypointDeformer (KD), SkeletonMerger (SM), SC3K on almost all categories on the ClothesNet dataset and ShapeNetCoreV2. Deformable objects keypoints seem to improve more. - Ablations in Tab. 4 indicate the importance on 4 (among dozens) of semantic classes how encoder-info, a grid heatmap, farthest point sampling loss, and similarity (i.e. chamfer loss) each boost performance.

Weaknesses

# Significance - Is the hierarchical decoder setup (excluding the heatmap grid itself) novel, esp compared to SkeletonMerger [26]? Right now there are no specific claims or ablations related to this. - The primary novelty relative to SM [26] seems to be the use of keypoint grid heatmaps in the decoder, storing max distances from grid cells to the weighted keypoint segments. However, this is just one possible shape descriptor feature, and intuitively it should be quite a weak and ad hoc feature, even if the authors gave some intuition why it may be better than the min function. Would other shape feature with yet more information be more useful could (e.g. why not have K channels with distances to all segments or to the keypoints as opposed to just taking the max, etc)? In general it would help if the paper provided more intuition or experimental results giving more insight on the information contained in this heatmap. - This class of approaches seems generally limited to cases where entire point clouds were visible. How would you handle cases when some parts of the objects were occluded, such as in real-world scenes? This has not beed discussed. # Related work Could discuss other dense representations for correspondence like the pointmaps used in “DUSt3R: Geometric 3D Vision Made Easy” https://arxiv.org/pdf/2312.14132 ) # Clarity - I found Sec 3.3 particularly difficult to follow and understand. A rewrite, or further details in the appendix clarifying better how these quantities are 'composed' and aligned would help. - Fig 1 illustrates points heatmap in the decoder by simply showing a point cloud. I was expecting pictures like in Fig 4c. - L13, L58 Unclear where keypoint pairs come from, when discussed for first time. - No discussion of how correspondences are established, e.g. as shown in Figure 3. - The loss terms in Equation 11 are not mathematically defined, would be helpful in the Appendix. # Experimental results - KeypointDeformer, SkeletonMerger, and SC3K are tested on KeypointNet dataset (https://arxiv.org/pdf/2002.12687 ). Why do the authors omit it here? Other works such as SC3K state “We use KeypointNet dataset [36 ] in our experiments, considering that this is the standard and most recent dataset used for keypoints estimation”. If the results are not as good on that dataset, that is okay, as long as we provide them and state that our method is specifically designed for datasets with deformable/soft objects. This would provide confidence that the authors ran their eval of official source code correctly on the two new datasets. - Why no comparisons with supervised baselines? - Figure 4D: Sun3D keypoints are so small in the figure it appears as if they are randomly distributed # Some Grammar & Style Nits: - L453 SC3K bibtex should be ICCV 2023, not Arxiv “ - L148 typo “giving a total of 4096 gird points.” -> “…4096 grid points” - L229 “mean Intersection over Unions” -> “mean Intersection over Union” - Grammar: Figure 3: caption: “Different methods on the Hat and Long Pant during” -> prefer “Different methods on the Hat and Long Pant categories during” - Table 3 caption is too close to Table 3, whitespace seems too shrunken - L75 “Deformable object dataset” -> “Deformable object datasets” - Table 1 caption “codes” -> should be singular, not plural (“the results we reproduced based on their official codes” -> “the results we reproduced based on their official code.”) - Color scheme is jarring in Tables 1-4, prefer tango colors that have a more attractive colormap (https://sobac.com/sobac/tangocolors.htm) - L3 extraneous article: “focus on the rigid body objects” -> “focus on rigid body objects” - L6 extraneous article: “for both the rigid-body and deformable objects” -> “for both rigid body and deformable objects” - L300 “the importance of encoder information and grid heatmap for reconstruction process,” -> “...for the reconstruction process,” - L302: “Table 4 show that the keypoints detected by Key-Grid which utilizes both two strategies to reconstruct the point cloud have better semantic consistency.” -> “Table 4 shows that when using both input streams to reconstruct the point cloud, the keypoints detected by Key-Grid exhibit better semantic consistency” - L9-10 Instead of “Unlike previous work, we leverage the identified keypoint information to form a 3D grid feature heatmap 1called grid heatmap, which is used in the decoder section”, prefer to say something like “..to form a 3D grid feature heatmap which we refer to as a grid heatmap. A grid heatmap is…” - L14 “Into the decoder section” -> “into the decoder model”? Ambiguous what “section” refers to - L51 “aiming at the semantic consistency” -> “aiming for semantic consistency”, e.g. “an unsupervised keypoint detector on 3D point clouds aiming for semantic consistency under shape variations of both rigid-body and deformable objects.” - L74 section should be named “Related Work” not “Related Works”

Questions

- Compared to SM[26], is the keypoint heatmap the main difference, how about the use of hierarchical decoder. - Wouldn't a richer shape descriptor heatmap do even better than a max function? (e.g. K channels, distances to the keypoints, etc). - How do you get keypoint correspondences across shapes for your model, do they emerge naturally? - What are the method results on KeypointNet dataset? Can you provide some additional comparisons to supervised baselines on your datasets, if there are such? - Can this method be applied on datasets where shapes are significantly occluded (e.g. viewed from a specific direction)?

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

Limitations have been sufficiently addressed.

Authorsrebuttal2024-08-12

Thank you for your detailed review and comments on our paper. We will revise our paper according to your suggestions. We are also pleased to hear that you believe our paper has reached the ’weak accept‘ stage. However, we observe that the rating is not updated. We would be grateful if you could revise the rating to better reflect the improvements made to the paper.

Reviewer p6XE2024-08-09

Additional comments

Thank you for the response and the additional experiments. My main concerns regarding the grid heatmap representation and ablation study have been resolved. The quality of the Figures 1 and 2 has also improved. I tend to accept this paper. However, I still have another question: can the distances between key points or the density of key points express the dense extension of the deformable object?

Authorsrebuttal2024-08-09

Thanks for your respone

In my view, using the distance from a point to keypoints to represent the deformable object’s structure is less effective than using the distance from a point to the skeleton. For example, in the main paper of Figure 2, if we use the distance from a point to keypoints, in the tail section of the pants (i.e., the area between the red and blue points), the value near the red point differs from the value at the center of this section. However, the geometric information represented by these two points should be consistent. By using the distance from a point to the skeleton, the point values are derived based on the skeleton spreading outward from the center. This results in different values for points near the skeletons compared to those at the object’s edges, reflecting their distinct geometric information. Therefore, we believe that using the distance from a point to the skeleton more accurately characterizes the structure of deformable objects than using the distance to keypoints.

Reviewer p6XE2024-08-10

Final Rating

Thank you for the response. The rebuttal has addressed my main concerns, and I am willing to raise my rating to "weakly accept."

Reviewer YtUf2024-08-10

rebuttal response

The authors have provided helpful answers and results that largely address my questions. I also saw the newly attached figure, which shows performance under some occlusion - which is ok even if the occlusion amount shown is quite small. In light of this, I am willing to raise my rating to 'weakly accept'.

Reviewer LZF72024-08-10

Final Rating

Thank you for the response. The rebuttal is very good and has addressed my major concerns, and I raise my rating to "weak accept."

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC