Temporal Action Localization by Structured Maximal Sums

TL;DR

This paper introduces a structured prediction approach for temporal action localization in videos, modeling actions as sums over temporal windows with explicit start, middle, and end components, solved efficiently with a new algorithm.

Contribution

It proposes a novel structured maximal sum algorithm for localizing actions, explicitly modeling temporal evolution with deep CNN features trained end-to-end.

Findings

Achieves competitive performance on THUMOS 14 benchmark

Develops a provably-efficient algorithm for structured prediction

Models temporal dependencies explicitly in action localization

Abstract

We address the problem of temporal action localization in videos. We pose action localization as a structured prediction over arbitrary-length temporal windows, where each window is scored as the sum of frame-wise classification scores. Additionally, our model classifies the start, middle, and end of each action as separate components, allowing our system to explicitly model each action's temporal evolution and take advantage of informative temporal dependencies present in this structure. In this framework, we localize actions by searching for the structured maximal sum, a problem for which we develop a novel, provably-efficient algorithmic solution. The frame-wise classification scores are computed using features from a deep Convolutional Neural Network (CNN), which are trained end-to-end to directly optimize for a novel structured objective. We evaluate our system on the THUMOS 14 action detection benchmark and achieve competitive performance.