Abstract

In this paper, we propose a novel encoding method for the representation of human action videos, that we call Discriminative Action Vector of Locally Aggregated Descriptors (DA-VLAD). DA-VLAD is motivated by the fact that there are many unnecessary and overlapping frames that cause non-discriminative codewords during the training process. DA-VLAD deals with this issue by extracting class-specific clusters and learning the discriminative power of these codewords in the form of informative weights. We use these discriminative action weights with standard VLAD encoding as a contribution of each codeword. DA-VLAD reduces the inter-class similarity efficiently by diminishing the effect of common codewords among multiple action classes during the encoding process. We present the effectiveness of DA-VLAD on two challenging action recognition datasets: UCF101 and HMDB51, improving the state-of-the-art with accuracies of 95.1% and 80.1% respectively.