ZOPP: A Framework of Zero-shot Offboard Panoptic Perception for Autonomous Driving

Offboard perception aims to automatically generate high-quality 3D labels for autonomous driving (AD) scenes. Existing offboard methods focus on 3D object detection with closed-set taxonomy and fail to match human-level recognition capability on the rapidly evolving perception tasks. Due to heavy reliance on human labels and the prevalence of data imbalance and sparsity, a unified framework for offboard auto-labeling various elements in AD scenes that meets the distinct needs of perception tasks is not being fully explored. In this paper, we propose a novel multi-modal Zero-shot Offboard Panoptic Perception (ZOPP) framework for autonomous driving scenes. ZOPP integrates the powerful zero-shot recognition capabilities of vision foundation models and 3D representations derived from point clouds. To the best of our knowledge, ZOPP represents a pioneering effort in the domain of multi-modal panoptic perception and auto labeling for autonomous driving scenes. We conduct comprehensive empirical studies and evaluations on Waymo open dataset to validate the proposed ZOPP on various perception tasks. To further explore the usability and extensibility of our proposed ZOPP, we also conduct experiments in downstream applications. The results further demonstrate the great potential of our ZOPP for real-world scenarios.

Paper

References (73)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 96Yj5/10 · confidence 3/52024-06-13

Summary

The paper introduces a novel framework, termed Multi-modal Zero-shot Offboard Panoptic Perception (ZOPP), specifically designed for autonomous driving applications. This innovative approach integrates zero-shot recognition capabilities with 3D representations generated from point cloud data, enhancing the model's ability to interpret complex driving environments without the need for extensive labeled datasets.

Strengths

1. The proposed work by the authors is significant and unique. 2. The proposed ZOPP approach is more capable of handling diverse types of data compared to any existing model, making it more robust.

Weaknesses

1. The manuscript is severely deficient in addressing the issues arising from labelling constraints and offers no substantive solutions within the proposed model to mitigate these problems. 2. The authors have egregiously failed to provide any details regarding the computational resources utilized, including the computational cost and the necessary hardware specifications, leaving the readers in the dark about the feasibility and scalability of the proposed approach. 3. The manuscript is conspicuously lacking in structure, offering an incoherent presentation of the proposed work, which undermines the clarity and comprehensibility of the research. 4. Despite the purported significance of the proposed work, the evaluation is lamentably limited, providing insufficient empirical evidence to substantiate the claimed advancements and benefits. 5. The manuscript is grossly inadequate in offering detailed information about the training and testing procedures, thereby failing to provide the essential methodological transparency required for replication and validation of the results.

Questions

1. The authors mention that data labeling is a crucial factor and a primary hurdle in training within existing work. Could the authors elaborate on how they propose to counter this issue or if they have developed any specific labeling mechanism to address this challenge? 2. How have the authors handled data from multi-view cameras? Specifically, how do they manage instances of the same object appearing in different views? 3. (With respect to Fig. 3), how does the proposed convolution filtering operation for removing background 3D points from foreground pixels account for varying disparity values at the upper edges of foreground objects? What specific kernel design is employed to optimize this process? 4. In the context of disparity occlusion handling around the upper edges of foreground objects (as seen in Fig. 3), what is the mathematical formulation for projecting 3D background points into the pixel regions of foreground objects? Additionally, how does the proposed convolution filtering algorithm differentiate between true foreground and occluded background points at a sub-pixel level? 5. What is the likelihood of the model's success in real-world applications? Have the authors conducted any real-world testing, and if so, what were the results?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

1. The abstract is grossly inadequate for the proposed topic. 2. The manuscript fails to deliver a coherent and sequential exposition of the proposed work. 3. The authors have entirely neglected to address the computational cost associated with the work or necessary for deployment. 4. The manuscript is deficient in presenting a comparative analysis with the visual results of existing works.

Reviewer Swqu4/10 · confidence 3/52024-07-05

Summary

Offboard perception creates 3D labels for autonomous driving scenes. Current methods are limited and don't match human recognition levels. The authors developed a new framework called Zero-shot Offboard Panoptic Perception (ZOPP), which combines advanced recognition technologies with 3D point cloud data. ZOPP is a pioneering approach in automatic labeling for driving scenes. They validated its effectiveness on the Waymo dataset and downstream applications with good performance.

Strengths

1. Using foundation model to generate labels for autonomous driving is meaningful task. it can help to push a great advance in the realm of autonomous driving. 2. Experiments are extensive with good results.

Weaknesses

1. the proposed work just combines some off-the-shelf modules together, and use some basic mathematical thing to support the alignment, which looks not very innovative as NeurIPS requested, although the validation results is competitive. 2. The presentation need to be improved. sometimes, it is hard to guess what you refer to.

Questions

refer to weaknesses

Rating

4

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

refer to weaknesses

Authorsrebuttal2024-08-13

We sincerely appreciate your efforts during the review and responses. We are wondering whether our response has addressed your concerns and do you have any more suggestions or questions. Moreover, if you find our response satisfactory, we kindly invite you to consider the possibility of improving your rating.

Reviewer aSzs5/10 · confidence 3/52024-07-12

Summary

ZOPP proposes an offboard auto-annotation method to achieve lidar 3D detection as well as the occupancy label without any annotation data. The whole pipeline ensembles several models including the SAM-track and point cloud completion model. By using some post-processing to complete the

Strengths

1. The idea is straightforward. 2. The writing is clear.

Weaknesses

Lack of novelty. It ensembles several SOTA methods but lacks in its contribution.

Questions

See weaknesses.

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

See weaknesses.

Authorsrebuttal2024-08-12

We sincerely appreciate your valuable and helpful suggestions. We are wondering whether you have any more suggestions or questions after our response. Specifically, do you have any questions about the performance comparison based on distances and occlusions? Moreover, if you find our response satisfactory, we kindly invite you to consider the possibility of improving your rating.

Reviewer aSzs2024-08-14

Thank you for the detailed explanation! It resolves my concerns on its performance.

Authorsrebuttal2024-08-14

We are grateful to you for recognizing our efforts to address your concerns during the response process. Your feedback has been instrumental in enhancing the quality of our work, especially for the comprehensive comparisons based on the distances and occlusions. We look forward to continuing to meet your expectations in the final version of our paper.

Reviewer c7Qa7/10 · confidence 3/52024-07-12

Summary

This paper introduces ZOPP, a framework for zero-shot panoptic perception of autonomous driving scenes. Leveraging image foundation models, ZOPP is able to perform zero-shot 3D object detection, 3D semantic segmentation, 3D panoptic segmentation, and 3D occupancy prediction, the first zero-shot model of its kind. Experiments on the Waymo dataset achieve strong performance.

Strengths

1. The paper is well-written and easy to follow. 2. ZOPP, to the best of my knowledge, is the first zero-shot panoptic perception model for autonomous driving constituting a significant novelty. 3. Experiments and ablation studies are extensive. 4. A successful zero-shot panoptic perception framework is highly useful for autolabeling driving scenes, giving this paper a high likelihood for significant impact.

Weaknesses

The most significant weakness is that zero-shot performance still significantly lags behind models trained using human-annotated labels. However, as this is the first work of it's kind, this is acceptable.

Questions

Do the authors have any quantitative results for their point completion module? This would be interesting to see.

Rating

7

Confidence

3

Soundness

4

Presentation

3

Contribution

4

Limitations

The authors adequately address limitations and broader impacts.

Reviewer 96Yj2024-08-07

In response to the feedback provided, the authors have agreed to revise the manuscript and address several of its current limitations. However, certain critical aspects of the manuscript still require closer examination and further refinement. Notably, significant limitations remain in areas such as the evaluation phase, and more results proof will be required, which strengthens the proposed model's useability.

Authorsrebuttal2024-08-14

We sincerely appreciate your efforts and valuable suggestions during the review and response periods. We are wondering whether you have any more specific questions or suggestions after our response. Your willingness to accept our paper is truly appreciated.

Reviewer Swqu2024-08-08

I have read the rebuttal

Reviewer c7Qa2024-08-12

I have read the author's rebuttal. Their response adequately addresses the questions I raised, and given the promise I believe this paper holds, I keep my rating as is.

Authorsrebuttal2024-08-14

We sincerely appreciate your thoughtful review and recognizing the ethical issues.

Authorsrebuttal2024-08-14

We sincerely appreciate your thoughtful review and recognizing the ethical issues. As we mentioned in Sec. 6 (Limitations and broader impacts) of our submission, we have discussed the potential concerns about data capturing and privacy. In the current public autonomous driving dataset, such as Waymo, nuScenes, KITTI, personal identifiable information of any person is masked, such as the human face, license plate of cars, and so on. Furthermore, we deeply agree with you to provide a section in the final version of our paper to discuss the concerns you have mentioned. We could provide several technical proposals to prevent misuse and illegal cases.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC