Decision Tree Based Early Termination Algorithm for Affine Prediction in AVS3

The third generation of Audio Video Standard(AVS3) has adopted affine prediction(including affine merge mode and affine inter mode), which can characterize non-translational motions such as rotation, zooming, and shearing. Though affine prediction leads to BD-Rate saving, it also results in substantial computational complexity. In this paper, we propose a fast algorithm to address this problem. Some useful features are introduced and used to train five decision tree classifiers to skip unnecessary calculations in affine prediction. Experimental results show that compared to HPM 14.0, the proposed algorithm reduces 19% affine merge time, 18% affine inter time, and 6% encoding time, with negligible coding performance loss under random access(RA) configurations.