USC & THU at THUMOS 2015

This notebook paper describes our approach for the action classification task of the THUMOS Challenge 2015. Our system combines motion and appearance features. For motion features, we adapt the Fisher vector representation with improved dense trajectories. For appearance feature, we compute feature activations from deep convolutional neural networks. We then train SVM classifiers and kernel ridge regression classifiers for each action class. During testing, these classifiers are applied to whole videos as well as temporal sliding windows with different durations. Finally, we combine the classification scores of the whole video and top ranked windows to generate video-level classification scores. Experimental results show that the proposed multi-duration fusion strategy improves the classification results significantly.