Model Checkpoints for VLPart

Object detection has been expanded from a limited number of categories to open vocabulary. Moving forward, a complete intelligent vision system requires understanding more fine-grained object descriptions, object parts. In this work, we propose a detector with the ability to predict both open-vocabulary objects and their part segmentation. This ability comes from two designs:

• We train the detector on the joint of part-level, object-level and image-level data.
• We parse the novel object into its parts by its dense semantic correspondence with the base object.