Unsupervised Action Anticipation Through Action Cluster Prediction

Predicting near-future human actions in videos has become a focal point of research, driven by applications such as human-helping robotics, collaborative AI services, and surveillance video analysis. However, the inherent challenge lies in deciphering the complex spatial-temporal dynamics inherent in typical video feeds. While existing works excel in constrained settings with fine-grained action ground-truth labels, the general unavailability of such labeling at the frame level poses a significant hurdle. In this paper, we present an innovative solution to anticipate future human actions without relying on any form of supervision. Our approach involves generating pseudo-labels for video frames through the clustering of frame-wise visual features. These pseudo-labels are then input into a temporal sequence modeling module that learns to predict future actions in terms of pseudo-labels. Apart from the action anticipation method, we propose an innovative evaluation scheme GreedyMapper, a unique many-to-one mapping scheme that provides a practical solution to the many-to-one mapping challenge, a task that existing mapping algorithms struggle to address. Through comprehensive experimentation conducted on demanding real-world cooking datasets, our unsupervised method demonstrates superior performance compared to weakly-supervised approaches by a significant margin on the 50Salads dataset. When applied to the Breakfast dataset, our approach yields strong performance compared to the baselines in an unsupervised setting and delivers competitive results to (weakly) supervised methods under a similar setting.