Rethinking the Faster R-CNN Architecture for Temporal Action Localization

Abstract

We propose TAL-Net, an improved approach to temporal action localization in video that is inspired by the Faster R-CNN object detection framework. TAL-Net addresses three key shortcomings of existing approaches: (1) we improve receptive field alignment using a multi-scale architecture that can accommodate extreme variation in action durations;(2) we better exploit the temporal context of actions for both proposal generation and action classification by appropriately extending receptive fields; and (3) we explicitly consider multi-stream feature fusion and demonstrate that fusing motion late is important. We achieve state-ofthe-art performance for both action proposal and localization on THUMOS'14 detection benchmark and competitive performance on ActivityNet challenge.

Learning Temporal Action Proposals With Fewer Labels

“…Buch et al [2] introduce semantics constraints for curriculum training in end-to-end temporal action localization. Chao et al [8] adopt Faster R-CNN [30] for action localization task.…”

Graph Convolutional Networks for Temporal Action Localization

“…Temporal Action localization has attracted increasing attention in the last several years [6,18,26,33,34]. Inspired by the success of object detection, most current action detection methods resort to the two-stage pipeline: they first generate a set of 1D temporal proposals and then perform classification and temporal boundary regression on each proposal individually.…”

Deep Point-Wise Prediction for Action Temporal Proposal

“…It combines 2D convolutional neural network and optical flow to capture appearance and motion features respectively. Recently, as kinds of 3D convolutional neural networks such as C3D [22], P3D [18], I3D [2] and 3D-ResNet [9] appear, adopting 3D convolutional neural network to extract spatio-temporal feature is getting more and more popular [1,2,25,3]. Temporal Action Proposals and Detection.…”

“…Temporal Action Proposals and Detection. Since natural videos are always long and untrimmed, temporal action proposals and detection have aroused intensive interest from researchers [6,26,1,25,3,8]. DAP [4] leverages LSTM to encode the video sequence for temporal features.…”

“…Other methods like TAL-net [3] modifies Faster-RCNN to fit temporal action proposal generation task. For temporal action detection, methods can be divided into two main types: one-stage [26,13,7,25,21] and two-stage [1,6,14].…”