Spatio-Temporal Instance Learning: Action Tubes from Class Supervision

The goal of this paper is spatio-temporal localization of human actions from their class labels only. The state-of-the-art casts the problem as Multiple Instance Learning, where the instances are a priori computed action proposals. Rather than disconnecting the localization from the learning, we propose a variant of Multiple Instance Learning that integrates the spatio-temporal localization during the learning. We make three contributions. First, we define model assumptions tailored to actions and propose a latent instance learning objective allowing for optimization at the box-level. Second, we propose a spatio-temporal box linking algorithm, exploiting box proposals from off-the-shelf person detectors, suitable for weakly-supervised learning. Third, we introduce tube- and video-level refinements at inference time to integrate long-term spatio-temporal action characteristics. Our experiments on three video datasets show the benefits of our contributions as well as its competitive results compared to state-of-the-art alternatives that localize actions from their class label only. Finally, our algorithm enables incorporating point and box supervision, allowing to benchmark, mix, and balance action localization performance versus annotation time.