Fast Collective Activity Recognition Under Weak Supervision.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Collective activity recognition is challenging due to complex human interactions.
• Existing methods often require extensive annotations (individual actions, poses, interactions) and decouple human detection, hindering practical application.
• Decoupled human detection increases computational load and struggles to filter irrelevant individuals.

Purpose of the Study:

• To develop a fast, weakly supervised deep learning architecture for collective activity recognition.
• To enable efficient and accurate group activity analysis by integrating actor detection and activity reasoning.
• To overcome limitations of previous methods requiring exhaustive annotations and separate detection steps.

Main Methods:

• An end-to-end framework that shares convolutional layers between actor detection and collective activity reasoning for fast inference.
• A joint learning approach that unifies and reinforces both tasks, enhancing the filtering of irrelevant individuals.
• A latent embedding scheme for mining person-group interactions, eliminating the need for pairwise relation labels and individual action labels.

Main Results:

• The proposed framework achieves comparable or superior performance to state-of-the-art methods on three benchmark datasets.
• The joint modeling approach enables collective activity reasoning at 22.65 frames per second, representing the fastest known speed.
• Demonstrates significant improvements in filtering out individuals not involved in the collective activity.

Conclusions:

• The developed weakly supervised deep learning architecture offers a fast and effective solution for collective activity recognition.
• The integrated approach of actor detection and activity reasoning significantly enhances efficiency and practical applicability.
• This research advances the field towards real-time collective activity recognition systems.