rpautrat
commented
9 months ago Hi, yes you can do your training on IR images with your existing keypoints. But you need to change a few things.
You need to convert your .txt files to the format we use in this repo: .npz files, one per image, which contains a dict {'points': np.ndarray}, where the ndarray is a num_keypoint x 2 array (in row-col convention, not x-y). These files must be present in a folder $EXPER_DIR/outputs/your-folder and later in your training config file you need to specify this dataset in the field data->labels.
Once this is done, you can just do step 6 to train directly the detector and descriptors.
And you may need to resize your images indeed. SuperPoint was trained on images 240x320. I recommend training on a similar size if possible. In that case, you need to also resize the keypoints accordingly.
Hello Rémi Pautrat Because I am a beginner in the Deep Learning field, I am so sorry if I bother you. I am facing some challenge with training my dataset, leading to that I really need your help for my problem: I have 10.000 RGB Images and 10.000 files containing keypoints which were detected by SuperPoint (x,y coordinates). My files are saved in “.txt” format. Now I have 10.000 IR Images which were aligned with 10.000 RGB Images. Now I want to using keypoints (on RGB images) with IR Images to train again for IR images:
Thank you so much for reading.