arjunkarpur
commented
4 months ago The released model doesn't need any training - you should be able to download the weights, load the model from the weights, and perform inference directly. Could you share some examples from your data? It may be a matter of hyperparam tuning, I can try it out.
In general, you would use a very low match threshold for pose estimation tasks (and rely on RANSAC to filter inliers), and a high match threshold when you require few but precise matches.
I tried your method on my data and found that the matching was unsatisfactory, so I was wondering if your method requires training. Because I don't see the training code.