Guess Where? Actor-Supervision for Spatiotemporal Action Localization
This addresses the problem of reducing annotation costs for action localization in videos, making it more accessible, though it is incremental as it builds on weakly-supervised techniques.
The paper tackles spatiotemporal action localization in videos using a weakly-supervised approach that only requires video class labels, achieving state-of-the-art results on three datasets and competitive performance with some fully-supervised methods.
This paper addresses the problem of spatiotemporal localization of actions in videos. Compared to leading approaches, which all learn to localize based on carefully annotated boxes on training video frames, we adhere to a weakly-supervised solution that only requires a video class label. We introduce an actor-supervised architecture that exploits the inherent compositionality of actions in terms of actor transformations, to localize actions. We make two contributions. First, we propose actor proposals derived from a detector for human and non-human actors intended for images, which is linked over time by Siamese similarity matching to account for actor deformations. Second, we propose an actor-based attention mechanism that enables the localization of the actions from action class labels and actor proposals and is end-to-end trainable. Experiments on three human and non-human action datasets show actor supervision is state-of-the-art for weakly-supervised action localization and is even competitive to some fully-supervised alternatives.