search

bluedream1121 / REKD

[CVPR 2022] Official PyTorch implementation of "Official Self-Supervised Equivariant Learning for Oriented Keypoint Detection"
MIT License
60 stars 6 forks source link

How to reproduce results in Figure 4 #14

Open Vathys opened 6 months ago

Vathys commented 6 months ago

Hello,

I'm interested in trying to reproduce the results in Figure 4 (repeatability under synthetic rotations). In particular, can you give me more information on how much Gaussian noise was added after applying the rotation and cropping? Section 4.4 of the ORB paper mentions "Gaussian noise of 10," but this is very vague.

Thank you

ph-code-repo commented 3 months ago

@Vathys I have been looking at this too. Between this paper and another of their papers on rotational equivariance, they do this experiment in different ways. From the desrciption in both papers, they seem to take a random sample of 10 images from HPatches and apply rotations in either 1deg/10deg intervals. Then, for the original image and each incrementally rotated image, the @5px MMA/repeatability/Orientation Accuracy (plus/minus 15deg) is calculated and plotted. At least this is what I seem to have been able to workout from the papers. I don't know what the Gaussian noise of 10 part is, but you can see the colour distortion parameters applied in their code (data/dataset_utils.py)