Awesome

Interactive Segmentation as Gaussian Process Classification (CVPR2023 Highlight)

Minghao Zhou, Hong Wang, Qian Zhao, Yuexiang Li, Yawen Huang, Deyu Meng, Yefeng Zheng

[Paper] [Poster] [Video] [Slides] [Supp]

Update 2023/8/1

We have updated is_gp_model.py for a more memory-efficient implementation.

Usage

Please first set up the environment and prepare the training (SBD)/testing (GrabCut, Berkeley, SBD, DAVIS) datasets following RITM, and change the directories in config.yml.

Please run run.sh for training/evaluation. For training, the resnet50 weights pretrained on ImageNet is used. Please download the weights and change the corresponding directory in config.yml. For evaluation, you can directly test with our provided checkpoint in checkpoints/GPCIS_Resnet50.pth.

The core codes of the GPCIS model can be found in isegm/model/is_gp_model.py and isegm/model/is_gp_resnet50.py.

Overview of GPCIS