Rethinking the Faster R-CNN Architecture for Temporal Action Localization

TL;DR

TAL-Net is a novel temporal action localization framework that enhances receptive field alignment, exploits temporal context, and emphasizes multi-stream feature fusion, achieving state-of-the-art results on key benchmarks.

Contribution

It introduces TAL-Net, which improves upon Faster R-CNN for video action localization by addressing receptive field, context exploitation, and feature fusion.

Findings

Achieves state-of-the-art on THUMOS'14

Competitive results on ActivityNet

Demonstrates importance of late motion fusion

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-of-the-art performance for both action proposal and localization on THUMOS'14 detection benchmark and competitive performance on ActivityNet challenge.