Part Discovery

Part discovery in computer vision aims to automatically identify and segment meaningful constituent parts of objects within images or 3D models, often without explicit part-level annotations. Current research focuses on leveraging deep learning architectures, including Vision Transformers and capsule networks, along with techniques like contrastive learning and diffusion models, to achieve unsupervised or weakly supervised part discovery. These advancements improve the interpretability of models and benefit downstream tasks such as fine-grained classification, object manipulation in robotics, and 3D model generation, leading to more robust and semantically rich representations. The development of novel metrics for evaluating part discovery performance is also an active area of research.