Rethinking Matching-Based Few-Shot Action Recognition

Few-shot action recognition, i.e. recognizing new classes given only a few examples, benefits from incorporating temporal information. Prior work either encodes such information in the representation itself and learns classifiers at test time, or obtains frame-level features performs pairwise matching. We first evaluate number of matching-based approaches using spatio-temporal backbones, comparison missing literature, show that gap performance between simple baselines more complicated methods is significantly reduced. Inspired by this, we propose Chamfer++, non-temporal matching function achieves state-of-the-art results few-shot recognition. that, when starting features, our parameter-free interpretable approach can outperform all other classifier for one-shot recognition on three common datasets without stage. Project page: https://jbertrand89.github.io/matching-based-fsar