DeHOT: Reconstructing Pseudo-3D Scenes for Human-Object Contact Detection

The goal of human-object contact detection (HOT) is to identify the contact areas in human-object interactions. Currently, existing work mainly focuses on segmenting the human-object contact regions without considering the spatial relationships among the human, the object, and the background, which leads to poor segmentation performance. To address this issue, we propose a method that utilizes depth maps to accomplish the HOT task. Specifically, depth maps are used to reconstruct pseudo-3D information about the scene, thereby obtaining the relative spatial relationships between people, objects, and the background. Subsequently, the combination of depth information with the actual input can further assist in determining the image segmentation areas. It is worth noting that, to address the overfitting issue in the HOT dataset, a comprehensive data augmentation strategy is proposed, which significantly improves the model’s accuracy.