Transferable Interactiveness Knowledge for Human-Object Interaction Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human-object interaction (HOI) detection is crucial for understanding human actions with objects.
• Existing HOI detection methods face challenges with diverse category settings and generalizing across datasets.

Purpose of the Study:

• To develop a transferable interactiveness knowledge learner for HOI detection.
• To improve the accuracy and generalization capabilities of HOI detection models.
• To alleviate the performance gap caused by diverse HOI category settings.

Main Methods:

• Exploited an Interactiveness Network to learn general interactiveness knowledge from multiple HOI datasets.
• Implemented Non-Interaction Suppression before HOI classification during inference.
• Utilized a hierarchical approach, learning instance-level and body part-level interactivenesses.
• Proposed a consistency task to guide learning and extract deeper visual clues.

Main Results:

• The Interactiveness Network demonstrated transferable knowledge learning capabilities.
• The method effectively suppressed non-interactions, improving HOI classification.
• Outperformed state-of-the-art HOI detection methods on HICO-DET, V-COCO, and HAKE-HOI datasets.
• Verified the efficacy and flexibility of the proposed approach.

Conclusions:

• Learned interactiveness knowledge is generalizable across HOI datasets and improves detection performance.
• The Interactiveness Network can be integrated with existing HOI detection models.
• The proposed method offers a flexible and effective solution for enhancing HOI detection.