Bidirectional Edge-Based 3D Scene Graph Generation from Point Clouds

The study of 3D scene understanding is a crucial area in computer vision and has received much attention. While ignoring the relationships between objects, current research on 3D scene understanding mostly focuses on object-level knowledge. In recent years, 3D scene graph has become an effective tool for attaining greater comprehension and perception of the surroundings. However, the long-tailed distribution of the training data causes existing 3D scene graph prediction models to yield sub-optimal scene graphs. In this paper, we provide a straightforward but effective 3D scene graph prediction model. Experiments on the 3DSSG dataset demonstrate that the model described in this research when compared to the baseline model, can increase relationship prediction's effectiveness and accuracy while successfully reducing the long-tailed distribution impact brought on by dataset labeling.