Highlighting Object Category Immunity for the Generalization of Human-Object Interaction Detection

Human-Object Interaction (HOI) detection plays a core role in activity understanding. As compositional learning problem (human-verb-object), studying its generalization matters. However, widely-used metric mean average precision (mAP) fails to model the well. Thus, we propose novel metric, mPD (mean Performance Degradation), as complementary of mAP evaluate performance gap among compositions different objects and same verb. Surprisingly, reveals that previous methods usually generalize poorly. With cue, Object Category (OC) Immunity boost HOI generalization. The idea is prevent from spurious object-verb correlations short-cut over-fit train set. To achieve OC-immunity, an OC-immune network decouples inputs OC, extracts representations, leverages uncertainty quantification unseen objects. In both conventional zero-shot experiments, our method achieves decent improvements. fully generalization, design new more difficult benchmark, on which present significant advantage. code available at https://github.com/Foruck/OC-Immunity.