Multi-Temporal-Resolution Technique for Action Recognition using C3D: Experimental Study

In any given video containing an action, the motion conveys information complementary to the individual frames. This motion varies in speed for similar actions. Therefore, it is a promising approach to train a separate deep-learning model for different versions of action speeds. In this paper, two novel ideas are explored: single-temporal-resolution single-model (STR-SM) and multi-temporal-resolution multi-model (MTR-MM). The STR-SM model is trained on one specific temporal resolution of the action dataset. This allows the model to accept a longer temporal frame range as input and therefore, a faster action classification. On the other hand, the MTR-MM is a set of STR-SM models, each trained on a different temporal resolution with a late fusion using majority voting achieving more accurate action recognition. Both models have improvements over the traditional training approach, 3.63% and 6% video-wise accuracy respectively.