search

rpautrat / SuperPoint

Efficient neural feature detector and descriptor
MIT License
1.86k stars 415 forks source link

Large difference in results between TF and pytorch #326

Open lyxxxyl opened 1 month ago

lyxxxyl commented 1 month ago

I have found when doing feature extraction on the same images that the tf version will all have a greater number of feature points than the pytorch version, the weights I use are sp_v6 and superpoint_v6_from_tf.pth respectively and when the input is a blank image, the tf result will be the points on the four corners of the image and the pytorch result will be empty.

lyxxxyl commented 1 month ago

When I do feature extraction on two images using the tf version, the shapes of keypoints, descriptors, scores are as follows image

And the result of torch version is as follows image

lyxxxyl commented 1 month ago

For images with many feature points, the difference is even greater image

rpautrat commented 1 month ago

Hi, this is probably due to different confidence threshold to decide how many keypoints you are keeping. Please check that you are using the same thresholds.

Note that there was already a similar issue on the topic: https://github.com/rpautrat/SuperPoint/issues/316