ACP++: Action Co-Occurrence Priors for Human-Object Interaction Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Human-object interaction (HOI) detection is crucial for understanding complex scenes.
• Training HOI detectors suffers from long-tailed distributions, with many rare classes having few labeled examples.
• This data imbalance leads to poor classification accuracy for infrequent HOI categories.

Purpose of the Study:

• To develop a novel approach for improving HOI detection accuracy, particularly for rare classes.
• To leverage natural correlations and anti-correlations between human actions and objects.
• To mitigate the negative impact of long-tailed data distributions in HOI datasets.

Main Methods:

• Modeling human-object interaction correlations using action co-occurrence matrices.
• Developing techniques to learn these interaction priors from data.
• Integrating learned priors into the training process for enhanced HOI detection.

Main Results:

• The proposed method demonstrates significant performance improvements over existing state-of-the-art techniques.
• Consistent gains in accuracy were observed across two major HOI benchmark datasets: HICO-Det and V-COCO.
• The approach effectively enhances the detection of rare human-object interaction classes.

Conclusions:

• Modeling interaction priors is an effective strategy for addressing data imbalance in HOI detection.
• The proposed action co-occurrence matrix approach offers a robust solution for improving rare class performance.
• This work advances the field of HOI detection by providing a more accurate and balanced model.