This repository contains the implementation of the following paper:
"Style Transfer for Keypoint Matching under Adverse Conditions".
Ali Uzpak, Abdelaziz Dejlouah, Simone Schaub-Meyer. 3DV 2020.
This code is based on D2-Net, R2D2 and Fast Style Transfer.
The code works with Python 3.6.10 and the following libraries:
- Pytorch 1.4.0 with CUDA 10.0.130 and cuDNN 7.6.3
- torchvision 0.5.0
- OpenCV 3.4.2
- scikit-image 0.17.2
- tqdm 4.41.1
- matplotlib 3.3.2
Newer versions of these libraries will very likely work as well.
Our machine is equipped with an Nvidia GTX 1080 Ti GPU that has 11GB of memory.
To match an image1 (night) to image2 (day), run
python optimize.py --image1_path input_images/dnim1.jpg --image2_path input_images/dnim2.jpg --num_steps_per_image 800 --content_steps 400 --output_path output_results/The execution time can be 3 to 5 minutes on a modern GPU.
The main parameters are:
image1_path: path to the image that has the adverse condition (night). This is going to be translated with the transformer.image2_path: path to the image that does not have the adverse condition (day).num_steps_per_image: total number of optimization steps for each image pair to be matched.content_steps: first number of optimization steps during which the transformer only minimizes the content loss.
For images im1 and im2, the results will be in output_path and consist of:
translated_im1_im2.png: translation (i.e. output of the transformer) of the night image1 according to the style of image2.original_matches_im1_im2_ninliers_X.png: image showing the matches (with #RANSAC inliers = X) using the D2-Net descriptors computed on the original image pair.new_matches_im1_im2_ninliers_Y.png: image showing the matches (with #RANSAC inliers = Y) using our method. D2-Net descriptors are computed on the translated image1 and matched with the D2-Net descriptors computed on the original image2.translator_data_im1_im2.npz: numpy file that encapsulates the R2D2 keypoints for both images, as well as the original and new matches (not filtered by RANSAC in this case). You can read this file with 'np.load()' and then access the keys 'kp1_r2d2', 'kp2_r2d2', 'translator_matches', 'original_d2net_matches'
Some precomputed results can be found in precomputed_output_results/.
The other examples of this repository can be computed with the following commands:
python optimize.py --image1_path input_images/dnim3.jpg --image2_path input_images/dnim4.jpg --num_steps_per_image 800 --content_steps 400 --output_path output_results/
python optimize.py --image1_path input_images/aachen1.jpg --image2_path input_images/aachen2.jpg --num_steps_per_image 800 --content_steps 400 --output_path output_results/- Input images are resized to make sure that they fit in the GPU memory. The allowed maximum size can be adapted to your GPU in the file
optimize.py. - R2D2 keypoints are combined with D2-Net descriptors using bilinear interpolation. This combination usually achieves the best results (cf. paper).
- Images in this repository come from the DNIM and Aachen Day-Night datasets.
To reproduce the paper's results on the DNIM dataset, we provide the dnim_matching_list.txt file which contains the 612 image pairs that we matched. However, if you want to match all those pairs using our optimization procedure, it will takes a while. We parallelized this task using 5 computers which took less than 1 day to finish.
With D2-Net descriptors only (124 inliers):
With our method (415 inliers!):
Similarily:
With D2-Net descriptors only (176 inliers):
With our method (472 inliers):
If you use this code in your project, please cite the following paper:
@inproceedings{Uzpak2020,
author={Uzpak, Ali and Djelouah, Abdelaziz and Schaub-Meyer, Simone},
title={Style Transfer for Keypoint Matching under Adverse Conditions},
booktitle = {3DV},
year = {2020}
}