Informative Sample Selection Model for Skeleton-Based Action Recognition With Limited Training Samples.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Skeleton-based human action recognition classifies skeletal data into action categories.
• Semi-supervised 3D action recognition addresses the challenge of limited annotated skeletal sequences.
• Active learning has been used to select informative samples for annotation in this domain.

Purpose of the Study:

• To improve semi-supervised 3D action recognition by developing a more effective active learning strategy.
• To address the limitation that representative samples are not always the most informative for model training.
• To reformulate the problem as a Markov Decision Process (MDP) for intelligent sample selection.

Main Methods:

• The study casts semi-supervised 3D action recognition via active learning as a Markov Decision Process (MDP).
• An informative sample selection model is trained within the MDP framework.
• Euclidean space factors are projected to hyperbolic space to enhance representational capacity.
• A meta-tuning strategy is introduced for faster real-world deployment.

Main Results:

• The proposed MDP-based active learning method effectively guides the selection of skeleton sequences for annotation.
• Experiments on three benchmarks demonstrate the method's effectiveness in improving 3D action recognition accuracy.
• The approach enhances the model's ability to learn from limited labeled data.

Conclusions:

• The novel MDP formulation provides a more intelligent way to select informative samples for semi-supervised 3D action recognition.
• Projecting to hyperbolic space and employing meta-tuning further boosts performance and applicability.
• This work offers a significant advancement in efficient and accurate 3D action recognition with limited annotations.