Collaborative and Multilevel Feature Selection Network for Action Recognition.

Area of Science:

• Computer Vision
• Deep Learning
• Action Recognition

Background:

• Feature pyramids are crucial for visual tasks but often treat features equally.
• Existing methods lack in-depth investigation into complementary advantages of different-level features.

Purpose of the Study:

• To propose a novel collaborative and multilevel feature selection network (FSNet) for robust action recognition.
• To adaptively aggregate multilevel features based on action context for enhanced representation.

Main Methods:

• FSNet employs position and channel selection modules with attention mechanisms.
• It adaptively aggregates multilevel features from position and channel dimensions.
• Features with different receptive fields and pattern-specific responses are emphasized and fused.

Main Results:

• FSNet demonstrated superior performance on Kinetics, UCF101, and HMDB51 datasets.
• The network can be flexibly inserted into various backbone networks.
• FSNet collaboratively trains to boost the representational ability of existing networks.

Conclusions:

• FSNet offers a practical approach to enhance action recognition models.
• The adaptive feature aggregation strategy significantly improves performance.
• FSNet achieves state-of-the-art results, surpassing existing top-tier models.