A Coarse-to-Fine Framework for Resource Efficient Video Recognition

摘要

Deep neural networks have demonstrated remarkable recognition results on video classification, however great improvements in accuracies come at the expense of large amounts of computational resources. In this paper, we introduce LiteEval for resource efficient video recognition. LiteEval is a coarse-to-fine framework that dynamically allocates computation on a per-video basis, and can be deployed in both online and offline settings. Operating by default on low-cost features that are computed with images at a coarse scale, LiteEval adaptively determines on-the-fly when to read in more discriminative yet computationally expensive features. This is achieved by the interactions of a coarse RNN and a fine RNN, together with a conditional gating module that automatically learns when to use more computation conditioned on incoming frames. We conduct extensive experiments on three large-scale video benchmarks, FCVID, ActivityNet and Kinetics, and demonstrate, among other things, that LiteEval offers impressive recognition performance while using significantly less computation for both online and offline settings.