LoFTR-MedicalData

Compare to the author's original work： LoFTR
Make weights/ folder using:

mkdir weights

Download the provided ckpts in the download link to

• 4 pretrained models of indoor-ds, indoor-ot, outdoor-ds and outdoor-ot (each ~45MB).

Finetune

Use the create_homo_data/Process_data.py to process each frame to get the folder 'xx_processed'

Set the soft links in data/homo/test by using:

ln -s train_processed/* data/homo/test/
ln -s val_processed/* data/homo_val/test/

After adding soft links:

Make the intrinsics.npz

• Using the create_homo_data/get_intrinsics.py to make the intrinsics.npz and homo_finetune.npz(they are same)
• put the intinsics.npz and homo_finetune.npz into assets/homo_fientune
• the content of homo_list_path.txt is the filename of homo_finetune.npz
  🚩 val_homo is also treated in the same way. Shown below:

The checkpoints you trained are in logs/

Test checkpoint

Input video to test

Change the input and model_ckpt of the run_demo.sh.
After that:

cd demo/
./run_demo.sh

Alignment results in demo/output:_ImagePairs
Matches are in demo/output:_MatchingPairs

Input image pairs to test

Change the img0_pth and img1_pth to the path of the image pair to be aligned in the ptb_colab2.py

python ptb_colab2.py