R2O PyTorch

PyTorch implementation of R2O from "Refine and Represent: Region-to-Object Representation Learning" (Gokul et al., 2022).

Pretrained Weights

We provide R2O ResNet-50 weights pretrained on ImageNet-1K for 300 epochs:

Format	Download	Use Case
Original	r2o_resnet50_imagenet300.pth	Direct loading
Torchvision	r2o_resnet50_imagenet300_torchvision.pth	MMSegmentation
Detectron2	r2o_resnet50_imagenet300_d2.pkl	Detectron2

Usage

See GitHub repo for how to use weights.

Citing this work

@misc{gokul2022refine,
  title = {Refine and Represent: Region-to-Object Representation Learning},
  author = {Gokul, Akash and Kallidromitis, Konstantinos and Li, Shufan and Kato, Yusuke and Kozuka, Kazuki and Darrell, Trevor and Reed, Colorado J},
  journal={arXiv preprint arXiv:2208.11821},
  year = {2022}
}