A Description of architectures 377 Motion Compensation We compare our method to the traditional motion-compensated coding

Quadratic Predictor We consider a multiscale architecture (cf. The top missing band of the images is cropped from 128 by 128 pixels to 112 by 128 pixels. Specifically, we trim a 17-pixel strip from each side, yielding frames of size 94 94 pixels. Figure 5: Filters of polar predictor networks trained to predict small synthetic sequences. In each panel, the 32 pairs of filters are sorted by their norm.