Magi-Net: Meta Negative Network for Early Activity Prediction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Early activity prediction aims to recognize actions from incomplete data.
• Partial video sequences often lack sufficient information for accurate classification, especially with few frames.
• Distinguishing similar activities from limited observations is a significant challenge.

Purpose of the Study:

• To propose a novel meta negative network (Magi-Net) for improved early activity prediction.
• To address the insufficiency of discriminative information in partial action sequences.
• To enhance the model's ability to learn from informative negative samples.

Main Methods:

• Developed Magi-Net, a network utilizing a contrastive learning scheme.
• Implemented a trainable negative look-up memory (LUM) table for negative sample selection.
• Introduced a meta negative sample optimization strategy (MetaSOS) for boosted training.

Main Results:

• Magi-Net demonstrated efficacy in early activity prediction tasks.
• The proposed model successfully alleviated the insufficiency of discriminative information.
• Experiments on public skeleton-based activity datasets validated the model's performance.

Conclusions:

• Magi-Net offers a promising solution for early activity prediction with limited observational data.
• The integration of contrastive learning and meta-learning strategies enhances classification accuracy.
• The approach is effective across various skeleton-based activity recognition benchmarks.