RESUMO
Recent methods including CoViAR and DMC-Net provide a new paradigm for action recognition since they are directly targeted at compressed videos (e.g., MPEG4 files). It avoids the cumbersome decoding procedure of traditional methods, and leverages the pre-encoded motion vectors and residuals in compressed videos to complete recognition efficiently. However, motion vectors and residuals are noisy, sparse and highly correlated information, which cannot be effectively exploited by plain and separated networks. To tackle these issues, we propose a joint feature optimization and fusion framework that better utilizes motion vectors and residuals in the following three aspects. (i) We model the feature optimization problem as a reconstruction process that represents features by a set of bases, and propose a joint feature optimization module that extracts bases in the both modalities. (ii) A low-rank non-local attention module, which combines the non-local operation with the low-rank constraint, is proposed to tackle the noise and sparsity problem during the feature reconstruction process. (iii) A lightweight feature fusion module and a self-adaptive knowledge distillation method are introduced, which use motion vectors and residuals to generate predictions similar to those from networks with optical flows. With these proposed components embedded in a baseline network, the proposed network not only achieves the state-of-the-art performance on HMDB-51 and UCF-101, but also maintains its advantage in computational complexity.
