Dynamic4D: Enhancing Self-Supervised Learning for Robust and Fine-Grained 4D Point Cloud Video Understanding

The proliferation of 4D point cloud videos highlights their potential, but the high cost of obtaining large-scale annotated data severely limits supervised methods. Consequently, self-supervised learning (SSL) is vital for learning generalizable representations from unlabeled 4D data. While existing SSL frameworks, such as Uni4D, have made progress, they often struggle with fine-grained motion understanding in extremely dynamic scenes, maintaining robustness under severe occlusion, and developing explicit predictive capabilities. To address these, we propose Dynamic4D, a novel and robust self-supervised framework tailored for dynamic 4D point cloud understanding. Dynamic4D introduces an Adaptive Causal Temporal Attention (ACTA) mechanism in the encoder for explicit causal temporal modeling and dynamic region-focused learning. Its decoder employs Motion Prediction Tokens (MPT) to directly infer motion vectors for masked regions. A novel adaptive motion-sensitive masking strategy further enhances robustness by intelligently prioritizing high-dynamic zones. Our multi-objective pre-training strategy integrates a new Dynamic Perception Loss alongside geometric reconstruction and latent-space alignment. Extensive experiments on diverse challenging benchmarks demonstrate that Dynamic4D consistently achieves state-of-the-art performance. It substantially outperforms prior methods, validating its superior capacity to learn highly robust, generalizable, and motion-aware representations for complex dynamic 4D point cloud scenes.