search

facebookresearch / silk

SiLK (Simple Learned Keypoint) is a self-supervised deep learning keypoint model.
GNU General Public License v3.0
644 stars 58 forks source link

Visualization #43

Closed zhangsngood closed 11 months ago

zhangsngood commented 1 year ago

Hi, when testing, I was hoping to get a visualization similar to the Figure 1 in your article. silk_matching

zhangsngood commented 1 year ago

But when I run the test code, I get the following visualization of the matches. I found that the color of the matching line in this way is not conducive to interpreting whether the matching result is good or bad. Can you please tell me how I can show the result like Figure 1. out_image

gleize commented 1 year ago

Hi @zhangsngood,

The visualization produced in the paper uses that function. In order to get the green / red colors for correct / incorrect matches, you need to provide ground truth matches (good_matches_mask argument). Since the inference code we provide doesn't have any ground truth (you can use any pair of images), we do show different colors in order to "manually" check if the correspondences are correct or not.

Looking at your image, the correspondences seems to look good, false positive matches would visually result in a lot of line crossing. To be certain of the quality, you would have to compute some metrics, using ground truth data (mean matching accuracy, homography estimation accuracy) or looking at the reconstruction statistics (number of registered images, re-projection error, etc).

gleize commented 11 months ago

Inactive for the last three weeks. Closing now.