COLMAP covisibility

rec_stats.py

def compute_reconstruction_statistics(reference_model_path):
    # Images w. intrinsics and extrinsics.
    with open(os.path.join(reference_model_path, 'cameras.txt'), 'r') as f:
        raw_cameras = f.readlines()[3 :]

    cameras = {}
    for raw_line in raw_cameras:
        split_line = raw_line.strip('\n').split(' ')
        cameras[int(split_line[0])] = split_line[1 :]

    with open(os.path.join(reference_model_path, 'images.txt'), 'r') as f:
        raw_images = f.readlines()[4 :]

    images = {}
    poses = {}
    intrinsics = {}
    for raw_line in raw_images[:: 2]:
        raw_line = raw_line.strip('\n').split(' ')
        image_path = raw_line[-1]
        image_name = image_path.split('/')[-1]
        image_id = int(raw_line[0])
        camera_id = int(raw_line[-2])
        intrinsics[image_path] = cameras[camera_id]
        images[image_path] = image_id
        poses[image_path] = parse_raw_pose(raw_line[1 : -2])

    # Covisibility matrix.
    image_visible_points3D = {}
    max_image_id = 0
    with open(os.path.join(reference_model_path, 'images.txt')) as images_file:
        lines = images_file.readlines()
        lines = lines[4 :]  # Skip the header.
        raw_poses = [line.strip('\n').split(' ') for line in lines[:: 2]]
        raw_points = [line.strip('\n').split(' ') for line in lines[1 :: 2]]
        for raw_pose, raw_pts in zip(raw_poses, raw_points):
            # image_id, qw, qx, qy, qz, tx, ty, tz, camera_id, name
            # points2D[(x, y, point3D_id)]
            image_id = int(raw_pose[0])
            max_image_id = max(max_image_id, image_id)
            point3D_ids = map(int, raw_pts[2 :: 3])
            image_visible_points3D[image_id] = set()
            for point3D_id in point3D_ids:
                if point3D_id == -1:
                    continue
                image_visible_points3D[image_id].add(point3D_id)

    n_covisible_points = np.zeros([max_image_id + 1, max_image_id + 1])
    # Fill upper triangle.
    for image_id1 in image_visible_points3D.keys():
        for image_id2 in image_visible_points3D.keys():
            if image_id1 > image_id2:
                continue
            visible_points3D1 = image_visible_points3D[image_id1]
            visible_points3D2 = image_visible_points3D[image_id2]
            n_covisible_points[image_id1, image_id2] = len(visible_points3D1 & visible_points3D2)
            # Mirror to lower triangle.
            n_covisible_points[image_id2, image_id1] = n_covisible_points[image_id1, image_id2]

    return images, intrinsics, poses, n_covisible_points