Demo on an image pair request

[fb82]

Hi,

could you provide a small demo to just run the feature matching on a single image pair and provide the correspondences?

Thanks in advance!

[Tangshitao]

You can refer to https://www.kaggle.com/code/optimo/2d-image-matching-with-quadtreeattention

[fb82]

Thanks!

Here my simple script based on the info in the link, that compute the matches and put them in a .mat file. Maybe can help someone...

import os
import argparse
import torch
import cv2
import numpy as np
import scipy.io as sio

from src.config.default import get_cfg_defaults
from src.utils.profiler import build_profiler
from src.utils.misc import lower_config

from src.loftr import LoFTR
from src.loftr.utils.supervision import compute_supervision_coarse, compute_supervision_fine

if __name__ == '__main__':
# run QuadTreeAttention setup.py in the other folder first !!!
    parser = argparse.ArgumentParser(
        description='QuadTreeAttention demo',
        formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    parser.add_argument('--weight', type=str, default="weights/outdoor.ckpt", help="Path to the checkpoint.")
    parser.add_argument('--config_path', type=str, default="configs/loftr/outdoor/loftr_ds_quadtree.py", help="Path to the config.")
    parser.add_argument('--input_1st', type=str, help="1st image.")
    parser.add_argument('--input_2nd', type=str, help="2nd image.")
    parser.add_argument('--output_file', type=str, default='quadtreeattention.mat', help=".mat file to save matches.")

    opt = parser.parse_args()

    # init default-cfg and merge it with the main- and data-cfg
    config = get_cfg_defaults()
    config.merge_from_file(opt.config_path)
    _config = lower_config(config)

    # Matcher: LoFTR
    matcher = LoFTR(config=_config['loftr'])
    state_dict = torch.load(opt.weight, map_location='cpu')['state_dict']
    matcher.load_state_dict(state_dict, strict=True)

    # Load example images
    img0_pth = opt.input_1st
    img1_pth = opt.input_2nd
    img0_raw = cv2.imread(img0_pth, cv2.IMREAD_GRAYSCALE)
    img1_raw = cv2.imread(img1_pth, cv2.IMREAD_GRAYSCALE)
    
   # img0_raw = cv2.resize(img0_raw, (640, 480))
   # img1_raw = cv2.resize(img1_raw, (640, 480))

    img0 = torch.from_numpy(img0_raw)[None][None].cuda() / 255.
    img1 = torch.from_numpy(img1_raw)[None][None].cuda() / 255.
    batch = {'image0': img0, 'image1': img1}

    # Inference with LoFTR and get prediction
    with torch.no_grad():
        matcher.eval()
        matcher.to('cuda')
        matcher(batch)
        mkpts0 = batch['mkpts0_f'].cpu().numpy()
        mkpts1 = batch['mkpts1_f'].cpu().numpy()
        mconf = batch['mconf'].cpu().numpy()
        mask_conf = mconf > 0
        conf = mconf[mask_conf]

    dict_to_save = {}        
    dict_to_save['kpt1'] = mkpts0
    dict_to_save['kpt2'] = mkpts1
    dict_to_save['conf'] = mconf
    
    sio.savemat(opt.output_file,dict_to_save)

[Tangshitao]

Thank you very much! I can commit this change if you send a pull request.

[fb82]

Just copy the code above in a file inside the FeatureMatching folder (I'm not good with git :| )...