Learnable Query Contrast and Spatio-temporal Prediction on Point Cloud Video Pre-training

Point cloud videos capture the time-varying environment and are widely used for dynamic scene understanding. Existing methods develop effective networks for point cloud videos but do not fully utilize the prior information uncovered during pre-training. Furthermore, relying on a single supervised task with a large amount of manually labeled data may be insufficient to capture the foundational structures in point cloud videos. In this paper, we propose a pre-training framework Query-CP to learn the representations of point cloud videos through multiple self-supervised pretext tasks. First, tokenlevel contrast is developed to predict future features under the guidance of historical information. Using a position-guided autoregressor with learnable queries, the predictions are directly contrasted with corresponding targets in the high-level feature space to capture fine-grained semantics. Second, performing only contrastive learning fails to fully explore the complementary structures and dynamics information. To alleviate this, a decoupled spatio-temporal prediction task is designed, where we use a spatial branch to predict low-level features and a temporal branch to predict timestamps of the target sequence explicitly. By combining the above self-supervised tasks, multi-level information is captured during the pre-training stage. Finally, the encoder is fine-tuned and evaluated for action recognition and dynamic semantic segmentation on three datasets. The results demonstrate the effectiveness of our Query-CP. Especially, compared with the state-of-the-art methods, the fine-tuning accuracy on action recognition improves by 3.23% for 24-frame point cloud videos, and the mean accuracy increases by 4.21%.