Disco4D: Towards Prior-Free Motion Representation Learning for Point Cloud Videos Via Self-Disentangled Contrastive Pre-Training

Point cloud video understanding requires joint modeling of geometry and motion. While MAE-based self-supervised frameworks allow geometry-aware supervision, they typically rely on hand-crafted or model-derived motion priors, which limits the learning of motion dynamics. In this work, we move beyond prior-dependent motion cues and learn motion from raw point cloud videos in a motion-prior-free manner. We propose Disco4D, a self-disentangled pre-training to extract expressive 4D representations. Our key insight is to separate motion learning in the high-level latent space of the MAE bottleneck while performing geometric reconstruction in the low-level Euclidean space. The motion supervision is achieved by aligning cross-frame features, including frame-level and video-level temporal dependencies. To address the representation gap caused by the entanglement of low-level and high-level features, we introduce two learnable tokens within a shared decoder. This design effectively disentangles low-level and high-level features and prevents overfitting. We show that our pre-trained encoder learns discriminative features without further fine-tuning. Extensive experiments on MSR-Action3D, NTU-RGBD, HOI4D, Synthia 4D, NvGesture, and SHREC’17 demonstrate the effectiveness of Disco4D for coarse-grained and fine-grained 4D downstream tasks. Moreover, we introduce a novel few-shot learning experiment to validate the generalization capability in 4D tasks.