AD-PT: Autonomous Driving Pre-Training with Large-scale Point Cloud Dataset

It is a long-term vision for Autonomous Driving (AD) community that the perception models can learn from a large-scale point cloud dataset, to obtain unified representations that can achieve promising results on different tasks or benchmarks. Previous works mainly focus on the self-supervised pre-training pipeline, meaning that they perform the pre-training and fine-tuning on the same benchmark, which is difficult to attain the performance scalability and cross-dataset application for the pre-training checkpoint. In this paper, for the first time, we are committed to building a large-scale pre-training point-cloud dataset with diverse data distribution, and meanwhile learning generalizable representations from such a diverse pre-training dataset. We formulate the point-cloud pre-training task as a semi-supervised problem, which leverages the few-shot labeled and massive unlabeled point-cloud data to generate the unified backbone representations that can be directly applied to many baseline models and benchmarks, decoupling the AD-related pre-training process and downstream fine-tuning task. During the period of backbone pre-training, by enhancing the scene- and instance-level distribution diversity and exploiting the backbone's ability to learn from unknown instances, we achieve significant performance gains on a series of downstream perception benchmarks including Waymo, nuScenes, and KITTI, under different baseline models like PV-RCNN++, SECOND, CenterPoint.

Paper

Similar papers

Peer review

Reviewer 5kGB6/10 · confidence 3/52023-07-01

Summary

This paper studies model pre-training for autonomous driving by utilizing a large-scale point cloud dataset, which largely facilitates the generalization performance of pre-trained models. To reconstruct such a dataset, scene- and instance-level distribution diversity are carefully enhanced. The overall pretraining is conducted in a semi-supervised manner, and its effectiveness is validated on several benchmark datasets with different baseline models.

Strengths

1. Compared to the traditional pre-training setting for point cloud, the AD-PT pretraining setting is more practical since it is expected to learn a generalized pre-trained model for all downstream tasks. 2. Compared to existing pre-trained methods trained in a fully unsupervised manner, this paper adopts a semi-supervised strategy, which is somewhat novel. 3. The data diversity enhancement with re-sampling strategies is reasonable. So does the unknown aware instance learning head.

Weaknesses

1. Although the AD-PT presented an improved performance on downstream tasks with limited training data (e.g., 20% data amount on Waymo), it is interesting to illustrate the results with more training samples in downstream tasks (e.g., 100% data amount on Waymo).

Questions

Na

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

NA

Reviewer nLKe5/10 · confidence 3/52023-07-01

Summary

This paper focuses on the point-cloud pre-training problem in the autonomous driving scenarios. Particularly, the authors regard the pre-training task as a semi-supervised learning problem and generate pseudo-labels for massive unlabeled data with few labeled frames. In the pseudo-label generation phase, they proposed two specific designs for this task, including diversity-based pre-training data preparation and unknown-aware instance learning. The experiments conducted on the KITTI, nuScenes, and Waymo datasets show the effectiveness of the proposed method.

Strengths

- The topic is meaningful. Different from the 2D vision task, the 3D detection community lacks effective pretraining methods and pre-trained backbones, and the strong pre-training backbones are helpful to the whole community. - The authors provide a pre-trained paradigm, and the experiments show the effectiveness of the pre-trained model, especially when training samples are limited. - Code will be publicly available. The authors provide source code for reproducibility.

Weaknesses

- The work claims that it provides a general pre-trained backbone for the LiDAR-based 3D object detection task. However, different from 2D perception tasks which are dominated by a few kinds of backbones, there are several popular detection pipelines with different backbone nets in this field. Some 3D detectors even use different input data representations, e.g. points, pillars, voxels, range images, etc. It is hard to build unified 3D pre-trained models for these works, and what this work provided is a voxel-based pre-trained backbone. Based on this, I think this work is over-claimed. - The 3D-based backbone is highly customized. Take the voxel-based backbone as an example, we need to pre-define the voxel size, and I think the pre-trained voxel size used in pretraining is the same as that used in the downstream task. I would like to know what would happen if different voxel sizes were used for pre-training and downstream tasks. If we must re-pretrain each backbone with different resolutions, then the significance of this work is not very important. - The improvement in performance is limited. I found that the pre-trained model has a relatively small impact on the final performance, especially in the full dataset setting. Considering the extra pre-training cost, performance improvement is limited, and I want to know whether simply extending the training schedule can achieve comparable performance or not. - Novelty is limited. This work regards the pre-training task as a semi-supervised learning problem with a pseudo-labeling scheme. However, using a few labeled data to generate pseudo-labels for LiDAR-based 3D detection has been investigated in [1], which achieves better performance. [1] Pseudo-labeling for Scalable 3D Object Detection, https://arxiv.org/abs/2103.02093

Questions

See Weaknesses.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

See Weaknesses.

Reviewer hHpC5/10 · confidence 4/52023-07-02

Summary

The paper proposes an autonomous driving pre-training method, where a large-scale pre-training dataset is built and a generalizable representation is created. To further improve the performances on the down-stream tasks, an advanced pre-training technique of semi-supervised method is proposed. Moreover, an unknown-aware instance learning head is proposed to learn the open-set detection. Various experiments are conducted to show the effectiveness of the method.

Strengths

1. The paper is well-organized and clearly written. 2. Thorough experiments are conducted on various benchmarks.

Weaknesses

1. The novelty is a bit limited, since the proposed techniques look like a combination of existing approaches, such as data augmentation and semi-supervised learning. 2. Although the improvements on various benchmarks are promising, the results are inferior compared to SOTA Lidar based object detection approaches such as VISTA [A] and MGTANet [B] on the nuScenes dataset. Hence, it would be great to also test on current SOTA approaches, which could further verify the generalization ability and robustness of the method. 3. Eq. 2 seems to be incorrect since x and y share the same calculation. It would be further clearer if a figure is drawn to show the calculation. 4. Line 293-294: Should the improvements for Waymo and nuScenes be 0.65%/0.54% and 7.37%, respectively? [A]: Deng, Shengheng, et al. "Vista: Boosting 3d object detection via dual cross-view spatial attention." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022. [B]: Koh, Junho, et al. "MGTANet: Encoding Sequential LiDAR Points Using Long Short-Term Motion-Guided Temporal Attention for 3D Object Detection." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 1. 2023.

Questions

None

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The limitation discussion seems to be missing. Possible memory consumption or time efficiency could be discussed to show a more comprehensive comparison with other SOTA works.,

Reviewer 8so55/10 · confidence 4/52023-07-06

Summary

This paper aims at large-scale point cloud dataset pre-training. The authors propose AD-PT method to build pre-training dataset with diverse data distribution and learn generalizable representations. In details, they design a diversity-based pre-training data preparation procedure and unknown-aware instance learning. The extensive experimental results on Waymo, nuScenes and KITTI datasets verify the effectiveness of the proposed method.

Strengths

1. Pre-training on the large-scale point cloud dataset is meaningful for autonomous driving. On the one hand, human annotation is expensive and how to use unlabeled data is an improtant direction. On the other hand, the point clouds of different LiDAR sensors have different patterns and we need consider these domain gaps. 2. The proposed method is reasonable. Although some thechniques are not new, suach as contrastive learning, I think they are reasonable for large-scale point cloud pretraining. For example, the LiDAR beam re-sampling can increase the robustness of beam domain gap. 3. The paper is well-written and easy to read.

Weaknesses

1. The results in Table 3 cannot show the superiority of the proposed method. The performance gain with SS-PT methods such as BEV-MAE is marginal. 2. I think a more useful setting is that we have some datasets A,B and C collected from different LiDAR sensors. We can use the labels in all three datasets and we want to improve the detector's performance on C dataset. Can the proposed method be applied to this setting? I hope the authors can give some insights. 3. The authors should give an analysis between the detector's performance on finetuning dataset and the labels amount in pre-training datasets. As an extreme case, if we do not have labels in pre-training dataset, will the performance drop largely?

Questions

Please reply to the weakness part.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

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 paper does not have limitation section.

Reviewer hHpC2023-08-15

Thanks for the rebuttal. To make the comparison to existing works fair enough, it is highly encouraged to directly compare to the widely used nuScenes testing data shown in Tab.1 of VISTA [A] and MGTANet [B] to show the effectiveness of the proposed method.

Authorsrebuttal2023-08-17

Thanks for your suggestion, and the nuScenes test results

We would like to thank the reviewer for this valuable suggestion. According to the reviewer's comment, we further submit the nuScenes test results of VISTA initialized by our proposed AD-PT to the **nuScenes server** to make a fair comparison. We follow VISTA to use double flip for testing augmentation. The test performance is shown below, where we use the performance of VISTA-OHS reported in their original paper for a fair comparison. | Methods | NDS | mAP | car | truck | cons. | bus | trailer | barrier | motorcycle | bicycle | pedestrian | traffic cone| | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | |VISTA-OHS (from scratch) | 69.8 | 63.0 | 84.4 | 55.1 | 25.1 | 63.7 | 54.2 | 71.4 | 70.0 | 45.4 | 82.8 | 78.5 | |VISTA-OHS (AD-PT) | 70.47 | 63.84 | 84.6 | 54.1 | 29.0 | 64.3 | 55.3 | 71.3 | 71.2 | 45.4 | 83.7 | 78.9 | We are very grateful for this insightful comment and will add the comparison of VISTA-OHS mentioned by the reviewer in the next version. Furthermore, we would like to emphasize that, the checkpoint obtained by our proposed AD-PT boosts the model performance not only on the nuScenes dataset (includes validation set and test set) but also on other public datasets, such as Waymo dataset (1.65% gain on PV-RCNN++ as shown in Table 3 of the main text), KITTI dataset (2.44% gain on PV-RCNN as shown in Table 5 of the main text).

Reviewer hHpC2023-08-19

Thanks for the reply and my concerns are addressed. I would like to change my rating to 'borderline accept'.

Reviewer nLKe2023-08-16

Final Rating

After reading the comments from other reviews and the rebuttal from the authors, I still think the novelty of this work is limited. However, I appreciate the additional experiments which shows the generalization ability of this work. So I'd like to slightly improve my score to 'borderline accept'

Authorsrebuttal2023-08-17

Thanks for your approval, and clarify our novelty

Thank you very much for your approval of AD-PT's generalization ability. We appreciate the reviewer's comments on our works. Due to the limited number of characters in the rebuttal phase, we would like to clarify the novelty of AD-PT from the following three aspects. **Motivation-level**: AD-PT is the first work aiming at using totally different pre-training and fine-tuning datasets and the first work that can provide backbones with strong generalization ability. Such a new pre-training paradigm can bring the following three advantages. - Data-efficient. Previous works mainly follow paradigms that use **the same dataset** for pre-training and fine-tuning. However, when the sensor is updated, the previous pre-training model can not improve the performance of the new data. In contrast, AD-PT can provide backbones with strong generalization ability, which means that the previous pre-trained backbone can improve the performance of new data. Meanwhile, in the rebuttal phase, we show that our method can easily combine with multi-dataset pre-training, which means we can use data from old sensors. **In conclusion, AD-PT can reuse previous data to provide a good pre-training model without re-collecting a large amount of data from the new sensor**. - Training-efficient. Previous methods need to pre-train new backbone parameters using new data when we want to improve the performance of newly collected data, which is time-consuming. **In contrast, AD-PT can achieve a one-shot pre-training to improve the performance of multiple domains, greatly reducing pre-training time and computation power consumption**. - Scaling up performance. Limited by the amount of data, previous works are hard to observe the performance gains when increasing the pre-training data. However, **AD-PT verifies that the downstream performance is positively correlated with the amount of the upstream pre-training data**. **Framework-level**: AD-PT is the first work to use semi-supervised pre-training and provides a complete set of pre-training processes including data labeling and training algorithm process (We will release the large-scale pseudo-labeled point clouds, about 1M data, which is the largest-scale labeled outdoor point-cloud dataset). Meanwhile, the starting point of AD-PT is totally different from previous semi-supervised learning, and data augmentation works as follows. - Previous pre-training methods mainly use MAE-based or contrastive learning-based methods to obtain backbone parameters to achieve the best in the single domain, while AD-PT explores a new pseudo-label-based method and totally considers the differences between upstream and downstream datasets, such as the differences in semantic classes between upstream and downstream. As a result, we propose to Unknown-aware Instance Learning Head to perform the open-set detection on the downstream dataset. - Furthermore, we provide a complete pre-training process including data preparation and a pre-training algorithm that is totally different from the previous pre-training methods as shown in Fig. 2. **Component-level**: Finally, each proposed module is also different from the existing ones. - Data preparation. Based on the observation that the quality of pseudo labels can affect the performance of pre-training process. We propose to use the class-aware pseudo generator and semi-supervised learning to obtain as accurate pseudo labels as possible. Pseudo-labeling as mentioned by the reviewer improves semi-supervised performance by using pseudo-labeling by means of threshold which is only a part of our method. Besides, our re-scaling and re-sampling methods meanly consider improving the diversity of pre-training data, which is different from most previous data augmentation methods aiming at improving the performance of the single-domain model. - Training algorithm. AD-PT is the first work that introduces the idea of open-set learning to the pre-training task, which fully considers the taxonomy difference between upstream and downstream datasets. We found some negative samples, that could be ignored during the upstream pre-training period, may be useful for different downstream fine-tuning tasks (due to the taxonomy differences between datasets). As a result, we consider such samples as undetected instances. Such an idea is different from previous MAE-based and contrastive-learning-based methods.

Reviewer 8so52023-08-20

Keep my positive rating

Thanks to the author's thoughtful response, I feel that my questions have been mostly resolved, and I will maintain my rather positive rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC