Reconstructing dynamic 3D scenes from a single video is difficult because motion must be inferred from observations that are noisy and only intermittently visible. RayCo4D, a new method described on arXiv, introduces a ray-conditioned continuous motion model to handle these challenges.
The approach leverages long-range tracks, which provide valuable geometric constraints, but their temporal information is fragmented. The model is designed to integrate these fragmented tracks into a continuous motion representation, improving the reconstruction of dynamic scenes.
The abstract does not provide quantitative results or implementation details, but it highlights the core problem of dealing with incomplete and noisy observations. The work appears to target monocular 4D reconstruction, where the goal is to recover both 3D structure and motion over time.