rpautrat
commented
4 years ago Hi,
What code are you using to get the keypoints? After inference with the model, are you directly using the 'pred' entry in the output? Or are you using the 'prob' (or 'prob_nms' if you want to apply NMS first) and then filtering it to keep the top_k points?
The 'pred' entry should indeed depend on the top_k and detection_threshold of the config file, but if the threshold is too high, you won't get enough keypoint (especially if the image sizes are small). I suggest doing the second option instead: use the 'prob_nms' entry and use a np.where(output['prob_nms'] > 0) to keep all the positive locations, and finally keep only the top_k best scores in it.
jiyun-xiao
Hi,
Thank you for your very well-written repo! I am trying to have around 1000 features when I do inference with the superpoint model. But I always get exactly 600 features when I do inference with the superpoint model I trained.
Here are the efforts I made to try to obtain more feature points:
I set the top_k = 2000 in config/superpoint_coco.yaml, but the super-point model that I exported still give exactly 600 points for any image that I input into it. (The detection threshold is the default value 0.001)
I thought that I was not getting enough features from magicpoint model, so I double checked that top_k was commented out in config/magic-point_coco_train.yaml and config/magic-poing_coco_export.yaml, but the magic-point model that I exported give exactly 1000 points for any image that I input into it.
I thus looked into the code for anywhere that might have caused the number of features to be exactly 1000 or exactly 600, but I haven’t found it.
I’m wondering if you could shine some light on what the problem might be? Thank you very much for your help!