Abstract:The area of temporally fine-grained video representation learning aims to generate frame-by-frame representations for temporally dense tasks. In this work, we advance the state-of-the-art for this area by re-examining the design of transformer architectures for video representation learning. A salient aspect of our self-supervised method is the improved integration of spatial information in the temporal pipeline by representing multiple entities per frame. Prior works use late fusion architectures that reduce frames to a single dimensional vector before any cross-frame information is shared, while our method represents each frame as a group of entities or tokens. Our Multi-entity Video Transformer (MV-Former) architecture achieves state-of-the-art results on multiple fine-grained video benchmarks. MV-Former leverages image features from self-supervised ViTs, and employs several strategies to maximize the utility of the extracted features while also avoiding the need to fine-tune the complex ViT backbone. This includes a Learnable Spatial Token Pooling strategy, which is used to identify and extract features for multiple salient regions per frame. Our experiments show that MV-Former not only outperforms previous self-supervised methods, but also surpasses some prior works that use additional supervision or training data. When combined with additional pre-training data from Kinetics-400, MV-Former achieves a further performance boost. The code for MV-Former is available at https://github.com/facebookresearch/video_rep_learning.