This paper presents CurveCodec 2, a skeletal animation codec that uses a small learned entropy model for its residuals.
Size: Against ACL, it needs 0.37x the bytes at 0.01 cm under a worst-case contract, and 0.22x at 0.1 cm under a mean-error contract, on 4,472 held-out clips from 33 datasets.
Speed and portability: It decodes on one CPU core, and its integer inference is bit-exact across platforms.
Unseen skeletons: It transfers to a species absent from training without retraining.
Where the learning helps: The authors say the network mainly learns the distribution of residuals. A nearest-neighbour oracle was no better than linear interpolation for the gaps, and no learned in-betweener they tried paid for itself.
Both error contracts are checked on every decoded clip.
Sign in to comment.
TL;DR of the paper, generated automatically.
This paper presents CurveCodec 2, a skeletal animation codec that uses a small learned entropy model for its residuals.
Both error contracts are checked on every decoded clip.