Temporal position information embedding method suitable for action recognition
Haowei Shi, Ying Hu · 2024
Temporal information is very important for action recognition, and it is still a challenge for video classification networks to better model temporal information. To better address this problem, this paper focuses on how to provide strong location information to action recognition networks so that the networks can better incorporate temporal information in understanding actions over long periods of time. In order to solve this problem, this paper proposes a way to encode the position of the input information for the action recognition network for video slice prediction, after the video slice for its position in the video is encoded and embedded into the feature map, which is sent to the network for prediction and fusion, in order to provide a strong temporal information, which is proved to be very effective, and our method is added to the TDN so that it still has a 2.58% improvement on the HMDB51 dataset.