Segment Any Point Cloud Sequences by Distilling Vision Foundation Models

Recent advancements in vision foundation models (VFMs) have opened up new possibilities for versatile and efficient visual perception. In this work, we introduce Seal, a novel framework that harnesses VFMs for segmenting diverse automotive point cloud sequences. Seal exhibits three appealing properties: i) Scalability: VFMs are directly distilled into point clouds, obviating the need for annotations in either 2D or 3D during pretraining. ii) Consistency: Spatial and temporal relationships are enforced at both the camera-to-LiDAR and point-to-segment regularization stages, facilitating cross-modal representation learning. iii) Generalizability: Seal enables knowledge transfer in an off-the-shelf manner to downstream tasks involving diverse point clouds, including those from real/synthetic, low/high-resolution, large/small-scale, and clean/corrupted datasets. Extensive experiments conducted on eleven different point cloud datasets showcase the effectiveness and superiority of Seal. Notably, Seal achieves a remarkable 45.0% mIoU on nuScenes after linear probing, surpassing random initialization by 36.9% mIoU and outperforming prior arts by 6.1% mIoU. Moreover, Seal demonstrates significant performance gains over existing methods across 20 different few-shot fine-tuning tasks on all eleven tested point cloud datasets.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 7qT56/10 · confidence 5/52023-07-03

Summary

This study represents the pioneering effort to utilize 2D vision foundation models for self-supervised representation learning on large-scale 3D point clouds. They introduce Seal, an innovative framework specifically designed to extract informative features from sequences of automotive point clouds. With its emphasis on scalability, consistency, and generalizability, Seal effectively captures semantic-aware spatial and temporal consistency, enabling the generation of highly valuable insights. Finally, they clearly demonstrate their superiority over previous state-of-the-art (SoTA) methods in both linear probing and fine-tuning for various downstream tasks.

Strengths

1. The paper is well-written and easily comprehensible. 2. To my knowledge, this is the first attempt to utilize large-scale vision models for aiding 3D point cloud segmentation. The authors have also put in tremendous effort, completing this work within just a short span of one to two months. 3. Conducted extensive experiments and achieved satisfactory results on various segmentation datasets.

Weaknesses

1. Strictly speaking, the approach in this paper is not considered unsupervised pretraining as it utilizes large models that rely on additional data. Most methods in Table 1, however, are unsupervised pretraining. Nevertheless, this distinction is not crucial, as the primary concern is to improve model performance, which is achieved through effective methods. 2. I believe that the approach presented in the paper has not fully distilled the knowledge of large 2D models. When using the complete dataset for both pretraining and fine-tuning, the information gain from distilling large models is not significant. Why do I say this? In the method employed in this paper, the large models primarily provide semantic supervision signals. If full semantic labels are given during fine-tuning, the information gain is limited. Although Table 2 shows significant improvements for some full dataset cases, this is mainly due to the use of semantic features in distillation rather than simple semantic pseudo-labels. The advantage of distilling large 2D models is that I can use more image-lidar pairs during pretraining. I believe that if additional image-lidar pairs are used, such as on the nuscenes dataset, there will be further improvements. I also hope to see similar experiments, such as pretraining with ten times the number of image-lidar pairs compared to nuscenes. 3. The method proposed in this paper is expected to be very general, so it should yield better results for other 3D perception tasks (e.g., 3D object detection).

Questions

The questions are all in the weaknesses section.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Limitations are all in the weaknesses section.

Reviewer 1xK97/10 · confidence 4/52023-07-06

Summary

This paper proposed a novel framework named Seal. The Seal distills VFMs into point clouds, enabling achieves efficient segmentation of various automotive point cloud sequences without requiring extensive annotation during the pre-training stage. It exhibits excellent performance across multiple datasets.

Strengths

The Seal distills the feature extraction capability of visual VFMs to point clouds. This paper promotes cross-modal representation learning through the proposed spatiotemporal consistency compared to previous works. The results obtained by The Seal on 11 different point cloud datasets demonstrate its effectiveness and superiority, highlighting its significant potential for 3D feature learning.

Weaknesses

The innovative contributions of the paper can be summarized into two points compared to previous works: 1. Distilling the representation learning capability from VFMs into point cloud data processing. 2. Introducing semantic superpoint temporal consistency to promote cross-modal learning. However, it is noted that the level of innovation in the paper might not be particularly strong.

Questions

1. I have some questions regarding the semantic superpoint temporal consistency. Firstly, according to Figure 4, the segmentation results provided by VFMs are at the instance level, while this paper aims to obtain a pretraining model for semantic segmentation tasks. According to Equation 2, this paper treats instances of the same object as positive samples and the rest as negative samples. Therefore, there is a high possibility that negative samples may include instances of the same class. This could potentially hurt the segmentation task. On the other hand, how is the correspondence between moving objects in point cloud data at different time steps obtained? 2. The paper mentions the ability to achieve good performance even in cases where camera and LiDAR calibration information is unavailable (on page 6 line 199). However, without approximate intrinsic and extrinsic parameters, establishing the 2D-3D spatial correspondence would be challenging. Therefore, I am curious about the performance in such scenarios. However, it appears that the paper does not provide specific experiments or results for these particular cases.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The statement in the limitation section that the assumption of obtaining calibrated and synchronized data between LiDAR and cameras is overly idealistic contradicts the description in the main text.

Reviewer 3rsk8/10 · confidence 4/52023-07-06

Summary

This paper introduces a novel framework, called *Seal*, that leverages VFMs for self-supervised representation learning on automotive point cloud sequences. The main idea is to leverage the 2D-3D correspondence between LiDAR and camera sensors and construct high-quality contrastive samples for cross-modal representation learning. The paper also proposes a superpoint temporal consistency regularization to enforce geometric stability across different timestamps. The paper evaluates the proposed framework on 11 different point cloud datasets and shows that it outperforms previous state-of-the-art methods in both linear probing and few-shot fine-tuning settings.

Strengths

**Originality:** - the **first** work utilizing 2D VFMs for self-supervised representation learning on large-scale 3D point clouds - a novel VFM-assisted contrastive learning objective that transfers the knowledge from the pretrained 2D to the 3D network at the semantic superpixel level - a superpoint temporal consistency regularization **Quality:** - well-written and organized, with clear motivation, problem formulation, related work, methodology, experiments, conclusion and discussion & limitations - extensive experiments on 11 different point cloud datasets and sufficient ablation studies - sufficient mathematical formulations and derivations to support the proposed methods - well-written and sufficient Supplementary Material with demo visualizations **Clarity:** - clear and easy to follow - uses consistent notations and symbols throughout the text and equations - helpful figures and tables to illustrate the main ideas and results **Significance:** - significant for both research and practice in 3D perception and representation learning - addresses an important problem of segmenting diverse automotive point clouds in a scalable, consistent, and generalizable manner

Weaknesses

The datasets conducted on are all **outdoor** point cloud dataset for autonomous driving. Unless the authors conduct more experiments on **indoor** point cloud datasets, they'd better just state that "Segment Any **Automotive** Point Cloud Sequences ..." in their title. Minior Issue: - Undefined variable $L^{tmp}$ in Equ. (3)

Questions

In Table 3 (Full), it seems *Seal* performs not well in `Cross`, `Echo` and `Sensor` (especially `Sensor`, 39.85 vs 49.21). Could you please provide some possible reasons about that?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

Authors should be rewarded that the limitations are explicitly mentioned in *Discussion & Limitation*.

Reviewer EPuy6/10 · confidence 5/52023-07-07

Summary

The manuscript presents Seal, a framework that leverages VFMs (Visual Foundation Models) to segment diverse point cloud sequences in autonomous driving scenarios. The proposed approach employs VFMs to initially segment superpixels in 2D camera images and subsequently projects them to 3D superpoints. Then, it incorporates two specific pre-training schemes: spatial contrastive learning, which involves training cross-modally from pre-trained 2D features to 3D representations, and temporal consistency regularization, which ensures the consistency of 3D point features across different timestamps. Extensive experimentation validates the effectiveness and robustness of the proposed framework.

Strengths

1. The manuscript exhibits clear and concise writing, making it easy to comprehend. The mathematical explanations and figures are presented with high clarity, facilitating understanding. 2. Through extensive experiments, the study demonstrates the effectiveness and robustness of the proposed method, establishing its reliability. 3. The main contribution of the manuscript is the development of the framework that can serve as a valuable resource for future reference and practical implementation.

Weaknesses

Novelty: The novelty of the manuscript may be somewhat limited. While the authors claim that the main novelty lies in leveraging VFM models for superpixel set segmentation and introducing two specific pretraining schemes, it is worth noting that the use of VFM models for segmentation is already a widely discussed topic and is short of novelty. Additionally, the proposed pretraining schemes closely follow common cross-modal contrastive learning approaches. This may give the impression that the paper is riding on the popularity of VFM models without truly exploring their potential for feature-level distillation.

Questions

1. What is the parameterization of a superpixel? If a superpixel is selected as a representative pixel from an area with the same class in VFM outputs, how is the pixel specifically chosen? 2. What is the motivation behind enforcing the convergence of 3D point features to a single mean representation? In the 2D scenario, it is common for features at different pixels of the same object to be distinct, as they capture diverse parts and information about the object. Shouldn't this principle apply to the 3D case?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Besides the limitations the authors discuss in the manuscript, I've seen no additional specific limitations.

Authorsrebuttal2023-08-10

General Response

We sincerely thank all the reviewers for the time and effort devoted to this review. --- We are glad to see that the reviewers are acknowledging this work: - *"is significant for both research and practice in 3D perception and representation learning"* (Reviewer 3rsk); - *"can serve as a valuable resource for future reference and practical implementation"* (Reviewer EPuy); - *"demonstrates the effectiveness and superiority, highlighting its significant potential for 3D feature learning"* (Reviewer 1xK9); - *"conducted extensive experiments and achieved satisfactory results on various segmentation datasets"* (Reviewer 7qT5). --- We would like to re-emphasize the novelty and main contributions of this work: - To the best of our knowledge, this study represents the first attempt at utilizing 2D vision foundation models for self-supervised representation learning on large-scale 3D point clouds. - The introduce Seal framework is scalable, consistent, and generalizable. This approach is designed to capture semantic-aware spatial and temporal consistency, enabling the extraction of informative features from automotive point cloud sequences. - Our approach demonstrates clear superiority over previous state-of-the-art methods in both linear probing and fine-tuning for downstream tasks across 11 different point cloud datasets with diverse data configurations. --- We have revised our manuscript to include the following changes according to the reviewers’ insightful comments: - We have supplemented more details on the definition and generation of semantic superpixel and superpoint. - We have added some remarks on the choice and motivation behind each design. - We conducted a study on the effect of possible misalignment between the LiDAR and camera sensors. - We have extended the Seal framework to other 3D perception tasks. - We have added experiments on image-LiDAR pre-training using sweeps data in nuScenes. - We have polished the elaboration and clarified some typos and misunderstandings in the main submission. --- Last but not least, we thank the PCs, ACs, and all the reviewers again for devoting time and effort to this review.

Authorsrebuttal2023-08-18

Looking forward to discussion

Dear Reviewer EPuy, We sincerely thank you for devoting time to this review and providing valuable comments. --- Based on the reviewers' comments, we have revised our manuscript to include the following changes: - We have supplemented more details on the definition and generation of semantic superpixel and superpoint. - We have added some remarks on the choice and motivation behind each design. - We conducted a study on the effect of possible misalignment between the LiDAR and camera sensors. - We have extended the Seal framework to other 3D perception tasks. - We have polished the elaboration and clarified some typos and misunderstandings in the main submission. --- We will actively participate in the Author-Reviewer discussion session. Please don't hesitate to let us know of any additional comments on the manuscript or the changes. Best regards, The Authors

Authorsrebuttal2023-08-18

Looking forward to discussion

Dear Reviewer 3rsk, We sincerely thank you for devoting time to this review and providing valuable comments. --- Based on the reviewers' comments, we have revised our manuscript to include the following changes: - We have supplemented more details on the definition and generation of semantic superpixel and superpoint. - We have added some remarks on the choice and motivation behind each design. - We conducted a study on the effect of possible misalignment between the LiDAR and camera sensors. - We have extended the Seal framework to other 3D perception tasks. - We have polished the elaboration and clarified some typos and misunderstandings in the main submission. --- We will actively participate in the Author-Reviewer discussion session. Please don't hesitate to let us know of any additional comments on the manuscript or the changes. Best regards, The Authors

Reviewer 3rsk2023-08-18

Thanks for the rebuttal and revised manuscript. The authors have addressed all my concerns and questions.

Authorsrebuttal2023-08-19

Authors' Response to Reviewer 3rsk

We sincerely thank Reviewer 3rsk for the positive feedback provided and the time devoted to this review. Best regards, The Authors

Authorsrebuttal2023-08-18

Looking forward to discussion

Dear Reviewer 1xK9, We sincerely thank you for devoting time to this review and providing valuable comments. --- Based on the reviewers' comments, we have revised our manuscript to include the following changes: - We have supplemented more details on the definition and generation of semantic superpixel and superpoint. - We have added some remarks on the choice and motivation behind each design. - We conducted a study on the effect of possible misalignment between the LiDAR and camera sensors. - We have extended the Seal framework to other 3D perception tasks. - We have polished the elaboration and clarified some typos and misunderstandings in the main submission. --- We will actively participate in the Author-Reviewer discussion session. Please don't hesitate to let us know of any additional comments on the manuscript or the changes. Best regards, The Authors

Authorsrebuttal2023-08-18

Looking forward to discussion

Dear Reviewer 7qT5, We sincerely thank you for devoting time to this review and providing valuable comments. --- Based on the reviewers' comments, we have revised our manuscript to include the following changes: - We have supplemented more details on the definition and generation of semantic superpixel and superpoint. - We have added some remarks on the choice and motivation behind each design. - We conducted a study on the effect of possible misalignment between the LiDAR and camera sensors. - We have extended the Seal framework to other 3D perception tasks. - We have polished the elaboration and clarified some typos and misunderstandings in the main submission. --- We will actively participate in the Author-Reviewer discussion session. Please don't hesitate to let us know of any additional comments on the manuscript or the changes. Best regards, The Authors

Authorsrebuttal2023-08-19

Authors' Response to Reviewer 7qT5

We thank Reviewer 7qT5 for devoting time to this review and providing valuable comments. --- Regarding Q2 (copied below) in the previous comment: > *"I believe that if additional image-lidar pairs are used, there will be further improvements. I hope to see similar experiments, such as pretraining with ten times the number of image-lidar pairs compared to nuScenes."* **A:** Thanks for the suggestion. We now present the follow-up updates for this question. - We believe pretraining with much more image-LiDAR pairs is indeed desirable. We use the *sweeps data* of nuScenes for this kind of pretraining, which contains unannotated image-LiDAR pairs around nine times larger (a total of ~ 260k frames) than the used keyframe data (~ 29k frames). - We use the baseline SLidR for pretraining in this update and will include more results for Seal in the revision. The pretraining results on nuScenes are shown in the following table: | Method | LP | 1% | 5% | 10% | 25% | Full | |:-|:-:|:-:|:-:|:-:|:-:|:-:| | Random | 8.10 | 30.30 | 47.84 | 56.15 | 65.48 | 74.66 | | PointContrast | 21.90 | 32.50 | - | - | - | - | | DepthContrast | 22.10 | 31.70 | - | - | - | - | | PPKT | 35.90 | 37.80 | 53.74 | 60.25 | 67.14 | 74.52 | | SLidR | 38.80 | 38.30 | 52.49 | 59.84 | 66.91 | 74.79 | | **SLidR (w/ sweeps)** | 39.57 | 39.40 | 53.21 | 60.69 | 67.44 | 75.09 | | ST-SLidR | 40.48 | 40.75 | 54.69 | 60.75 | 67.70 | 75.14 | | Seal | 44.95 | 45.84 | 55.64 | 62.97 | 68.41 | 75.60 | - We observe that using more image-LiDAR pairs during pretraining indeed improves the performance, for both linear probing and downstream fine-tuning tasks with different ratios. - The improvements are not linearly scaled up with the amount of data used for pretraining. We conjecture that this is because the *sweeps data* of nuScenes are less diverse, since they were collected sequentially. - Nevertheless, we believe that pretraining with more image-LiDAR pairs is promising and desirable; with the advent of datasets with larger scales, the proposed pretraining framework could become more powerful and can achieve better cross-dataset performance. --- Last but not least, we thank Reviewer 7qT5 again for the time and effort devoted to this review.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC