Observation-Aligned Mask Priors for Learning Physical Dynamics from Authentic Occlusions

The Observation-Aligned Mask Priors framework tackles the issue of learning physical dynamics from incomplete observations, particularly when occlusions are structured and dependent on samples. Initially, a Bayesian Flow Network (BFN) is pretrained using binary observation masks to understand true occlusion structures. Subsequently, BFN sampling is directed by a globally normalized cross-entropy objective, producing sample-specific masks that correspond to each sparse observation. The context is established by the overlap of the guided and observed masks, while the remaining observed entries serve as query targets for a diffusion-based reconstruction model. This innovative method eliminates the need for heuristic masking rules or preset mask distributions, facilitating more precise dynamics learning from genuine occlusions.