Multi-Moments in Time: Learning and Interpreting Models for Multi-Action Video Understanding

TL;DR

This paper introduces the Multi-Moments in Time dataset with over two million labels for multi-action video recognition, addressing the challenge of multi-label annotations and demonstrating baseline models, visualization techniques, and transfer learning benefits.

Contribution

The paper presents a large-scale multi-label dataset for multi-action video understanding and develops baseline models and interpretability methods tailored for this complex task.

Findings

Baseline multi-action recognition results on M-MiT.

Improved visualization and interpretation techniques for multi-label models.

Transfer learning from M-MiT enhances performance on smaller datasets.

Abstract

Videos capture events that typically contain multiple sequential, and simultaneous, actions even in the span of only a few seconds. However, most large-scale datasets built to train models for action recognition in video only provide a single label per video. Consequently, models can be incorrectly penalized for classifying actions that exist in the videos but are not explicitly labeled and do not learn the full spectrum of information present in each video in training. Towards this goal, we present the Multi-Moments in Time dataset (M-MiT) which includes over two million action labels for over one million three second videos. This multi-label dataset introduces novel challenges on how to train and analyze models for multi-action detection. Here, we present baseline results for multi-action recognition using loss functions adapted for long tail multi-label learning, provide improved methods for visualizing and interpreting models trained for multi-label action detection and show the strength of transferring models trained on M-MiT to smaller datasets.