Block prediction using approximate template matching

Template matching methods have been shown to offer bit-rate savings of up to 15% when used for in-loop prediction in compression. Yet the required nearest-template search process results in prohibitive complexity. Hence, in this paper we use approximate nearest neighbor search methods to successfully address this drawback of template matching methods. Our approach uses a template index that is updated during the decoding process, yet the incurred overhead pays off in reduced nearest-template search complexity, resulting in a significant gain in template search complexity. Rate-distortion experiments further indicate that there is no rate-distortion penalty resulting from our proposed approximate template search method, and in fact a small gain of 0.1 dB is observed.