Pyramidal Signed Distance Learning for Spatio-Temporal Human Shape Completion

We address the problem of completing partial human shape observations as obtained with a depth camera. Existing methods that solve this can provide robustness, for instance model-based strategies rely on parametric models, or precision, learning approaches capture local geometric patterns using implicit neural representations. investigate how to combine both properties novel pyramidal spatio-temporal model. This model exploits signed distance fields in coarse-to-fine manner, order benefit from ability representations preserve geometry details while enforcing more global spatial consistency estimated shapes through features at coarser levels. In addition, our also leverages temporal redundancy integrate information over neighboring frames. Experiments standard datasets show and aggregation contribute outperform state-of-the-art completion.