Can AI understand 3D scenes as well as their individual parts?

Existing 3D AI models understand objects but miss fine-grained details—like distinguishing a chair's leg from its seat. PAR3D adds part-awareness by learning hierarchical object-part relationships and grounding both in 3D space. A new dataset (ScenePart) with part-level annotations enables training and evaluation. The method substantially improves part-focused tasks while maintaining strong performance on standard object-level benchmarks.