Large scale video action understanding

Abstract

The goal of the project is to build a large scale video dataset called Moments, and train existing/novel models for action recognition. To aid automation of video collection and annotation selection, I trained Convolutional Neural Network models to estimate the likelihood of a desired action appearing in video clips. Selecting clips, which are highly probable to contain the wanted action, for annotation leads to a more efficient process overall with higher yield. Once a sizable dataset had been amassed, I investigated new multi-modal models that make use of different (spatial, temporal, auditory) signals in the video. I also conducted preliminary experiments into several promising directions that Moments opens up, including multi-label training. Lastly, I trained baseline models on Moments to calibrate the performance of existing techniques. Post-training, I diagnosed the shortcomings of the models and visualized videos that were found to be particularly difficult. I discovered that the difficulty largely arises due to the great variety in quality/perspective/subjects found in Moments videos. This highlights the challenging nature of the dataset and its value to the research community.