Rb-PaStaNet: A Few-Shot Human-Object Interaction Detection Based on Rules and Part States

TL;DR

This paper introduces Rb-PaStaNet, a method that incorporates human prior knowledge through rules and part states to enhance the detection of rare human-object interactions, achieving notable improvements especially in rare classes.

Contribution

The paper proposes Rb-PaStaNet, integrating human-labeled rules into PaStaNet to specifically improve rare HOI class detection, addressing a key challenge in the field.

Findings

Improved detection accuracy for rare HOI classes.

Significant overall performance enhancement.

Better generalization to underrepresented interactions.

Abstract

Existing Human-Object Interaction (HOI) Detection approaches have achieved great progress on nonrare classes while rare HOI classes are still not well-detected. In this paper, we intend to apply human prior knowledge into the existing work. So we add human-labeled rules to PaStaNet and propose Rb-PaStaNet aimed at improving rare HOI classes detection. Our results show a certain improvement of the rare classes, while the non-rare classes and the overall improvement is more considerable.