To create the training environment, run
conda env create -f sph.yaml
Specify the dataset and log path in config/conf_file.yaml then run:
python train.py --config config/conf_file.yamlTo ensure exact comparison, our evaluation is based on that of sd-dino. First, create a conda environment following the sd-dino instructions, then run
python pck_spair_pascal_sphere.py --SAMPLE 0
Additional arguments that we introduced for our method:
--SPHto perform fused evaluation with a pretrained sphere mapper--KAPto compute Keypoint Average precision instead of PCK--DATA_PATHto specify the path to the evaluation set--SPH_CKPT_PATHto specify the path to a spherical mapper checkpoint
- sd-dino investigates unsupervised correspondences emerging from recent deep models that inspired this work
- geoaware-sc is a followup paper that comes to the same conclusion about geometry-related issues and proposes to fix them at test-time
If you find our work useful, please cite:
@InProceedings{mariotti2024improving,
author = {Mariotti, Octave and Mac Aodha, Oisin and Bilen, Hakan},
title = {Improving Semantic Correspondence with Viewpoint-Guided Spherical Maps},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2024},
pages = {19521-19530}
}