PdiscoFormer PartImageNet Seg Model (K=16)

PdiscoFormer (Vit-base-dinov2-reg4) trained on PartImageNet Seg with K (number of unsupervised parts to discover) set to a value of 16.

PdiscoFormer is a novel method for unsupervised part discovery using self-supervised Vision Transformers which achieves state-of-the-art results for this task, both qualitatively and quantitatively. The code can be found in the following repository: https://github.com/ananthu-aniraj/pdiscoformer

BibTex entry and citation info

@misc{aniraj2024pdiscoformerrelaxingdiscoveryconstraints,
      title={PDiscoFormer: Relaxing Part Discovery Constraints with Vision Transformers}, 
      author={Ananthu Aniraj and Cassio F. Dantas and Dino Ienco and Diego Marcos},
      year={2024},
      eprint={2407.04538},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.04538}, 
}