Untrimmed Video Classification for Activity Detection: submission to ActivityNet Challenge

TL;DR

This paper introduces a method for activity detection in untrimmed videos by combining global classification with frame-level analysis and dynamic programming, demonstrating that classification models can aid in temporal detection.

Contribution

The paper presents a novel approach that leverages untrimmed classification to improve activity detection in untrimmed videos, integrating global features with frame-level analysis.

Findings

Effective temporal activity proposals generated from untrimmed classification.

Combining global and frame-level analysis improves detection accuracy.

Method achieves competitive results in ActivityNet Challenge.

Abstract

Current state-of-the-art human activity recognition is focused on the classification of temporally trimmed videos in which only one action occurs per frame. We propose a simple, yet effective, method for the temporal detection of activities in temporally untrimmed videos with the help of untrimmed classification. Firstly, our model predicts the top k labels for each untrimmed video by analysing global video-level features. Secondly, frame-level binary classification is combined with dynamic programming to generate the temporally trimmed activity proposals. Finally, each proposal is assigned a label based on the global label, and scored with the score of the temporal activity proposal and the global score. Ultimately, we show that untrimmed video classification models can be used as stepping stone for temporal detection.