Multi-attention Networks for Temporal Localization of Video-level Labels

TL;DR

This paper introduces an ensemble of attention-based neural networks for temporal localization of video labels, effectively handling noisy data and improving accuracy in large-scale video understanding tasks.

Contribution

The work proposes a novel multi-attention network architecture with multiple attention mechanisms and ensemble strategies for segment-level video classification.

Findings

Achieved competitive ranking in the YouTube-8M challenge

Effectively handled noisy labels with attention mechanisms

Improved performance through model ensemble and fine-tuning

Abstract

Temporal localization remains an important challenge in video understanding. In this work, we present our solution to the 3rd YouTube-8M Video Understanding Challenge organized by Google Research. Participants were required to build a segment-level classifier using a large-scale training data set with noisy video-level labels and a relatively small-scale validation data set with accurate segment-level labels. We formulated the problem as a multiple instance multi-label learning and developed an attention-based mechanism to selectively emphasize the important frames by attention weights. The model performance is further improved by constructing multiple sets of attention networks. We further fine-tuned the model using the segment-level data set. Our final model consists of an ensemble of attention/multi-attention networks, deep bag of frames models, recurrent neural networks and convolutional neural networks. It ranked 13th on the private leader board and stands out for its efficient usage of resources.