Shift-tolerant Perceptual Similarity Metric
Abhijay Ghildyal, Feng Liu. In ECCV, 2022. [Arxiv]
Todo
Please run python lpips_2imgs.py
Please download the original BAPPS dataset using this script (here). Then, update path to the dataset in global_config.json.
To reproduce the results in the paper run the following:
AlexNet Vanilla
nohup bash n_pixel_shift_study/test_scripts/test.sh alex vanilla 2 64 50 > logs/eval_alex_vanilla.out &
AlexNet Shift-tolerant
nohup bash n_pixel_shift_study/test_scripts/test.sh alex shift_tolerant 1 64 50 > logs/eval_alex_shift_tolerant.out &
Note: To train and test our models in this paper, we used Image.BICUBIC. The results are similar when other resizing methods are used. Please feel free to switch back to bilinear as used in the original LPIPS work (here).
For other evaluations refer to ./n_pixel_shift_study/.
If you find this repository useful for your research, please use the following.
@inproceedings{ghildyal2022stlpips,
title={Shift-tolerant Perceptual Similarity Metric},
author={Ghildyal, Abhijay and Liu, Feng},
booktitle={ECCV},
year={2022}
}
This repository borrows from LPIPS, Anti-aliasedCNNs, and CNNsWithoutBorders. We thank the authors of these repositories for their incredible work and inspiration.