Abstract
We propose a new perspective to perform action recognition without temporal modeling. We cast the video recognition problem as an image recognition task, and show that an image classifier can suffice for video understanding with no bells and whistles. Our approach is extremely simple, and it composes input frames into a super image to train a classifier to fulfill the task of action recognition, in exactly the same way as classifying an image. We prove the viability of our idea by demonstrating strong and promising performance on three public datasets including Kinetics400, Moments and Jester, using a recently developed vision transformer. We also experiment with the prevalent ResNet image classifiers in vision to further validate our idea. Our approach achieves comparable results on Kinetics400 compared to some more sophisticated CNN approaches based on spatio-temporal modeling.
