utils.py

# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Shared utility functions for OmniGlue."""

import cv2
import numpy as np

from typing import Optional


def visualize_matches(
    image0: np.ndarray,
    image1: np.ndarray,
    kp0: np.ndarray,
    kp1: np.ndarray,
    match_matrix: np.ndarray,
    match_labels: Optional[np.ndarray] = None,
    show_keypoints: bool = False,
    highlight_unmatched: bool = False,
    title: Optional[str] = None,
    line_width: int = 1,
    circle_radius: int = 4,
    circle_thickness: int = 2,
    rng: Optional['np.random.Generator'] = None,
):
  """Generates visualization of keypoints and matches for two images.

  Stacks image0 and image1 horizontally. In case the two images have different
  heights, scales image1 (and its keypoints) to match image0's height. Note
  that keypoints must be in (x, y) format, NOT (row, col). If match_matrix
  includes unmatched dustbins, the dustbins will be removed before visualizing
  matches.

  Args:
    image0: (H, W, 3) array containing image0 contents.
    image1: (H, W, 3) array containing image1 contents.
    kp0: (N, 2) array where each row represents (x, y) coordinates of keypoints
      in image0.
    kp1: (M, 2) array, where each row represents (x, y) coordinates of keypoints
      in image1.
    match_matrix: (N, M) binary array, where values are non-zero for keypoint
      indices making up a match.
    match_labels: (N, M) binary array, where values are non-zero for keypoint
      indices making up a ground-truth match. When None, matches from
      'match_matrix' are colored randomly. Otherwise, matches from
      'match_matrix' are colored according to accuracy (compared to labels).
    show_keypoints: if True, all image0 and image1 keypoints (including
      unmatched ones) are visualized.
    highlight_unmatched: if True, highlights unmatched keypoints in blue.
    title: if not None, adds title text to top left of visualization.
    line_width: width of correspondence line, in pixels.
    circle_radius: radius of keypoint circles, if visualized.
    circle_thickness: thickness of keypoint circles, if visualized.
    rng: np random number generator to generate the line colors.

  Returns:
    Numpy array of image0 and image1 side-by-side, with lines between matches
    according to match_matrix. If show_keypoints is True, keypoints from both
    images are also visualized.
  """
  # initialize RNG
  if rng is None:
    rng = np.random.default_rng()

  # Make copy of input param that may be modified in this function.
  kp1 = np.copy(kp1)

  # Detect unmatched dustbins.
  has_unmatched_dustbins = (match_matrix.shape[0] == kp0.shape[0] + 1) and (
      match_matrix.shape[1] == kp1.shape[0] + 1
  )

  # If necessary, resize image1 so that the pair can be stacked horizontally.
  height0 = image0.shape[0]
  height1 = image1.shape[0]
  if height0 != height1:
    scale_factor = height0 / height1
    if scale_factor <= 1.0:
      interp_method = cv2.INTER_AREA
    else:
      interp_method = cv2.INTER_LINEAR
    new_dim1 = (int(image1.shape[1] * scale_factor), height0)
    image1 = cv2.resize(image1, new_dim1, interpolation=interp_method)
    kp1 *= scale_factor

  # Create side-by-side image and add lines for all matches.
  viz = cv2.hconcat([image0, image1])
  w0 = image0.shape[1]
  matches = np.argwhere(
      match_matrix[:-1, :-1] if has_unmatched_dustbins else match_matrix
  )
  for match in matches:
    pt0 = (int(kp0[match[0], 0]), int(kp0[match[0], 1]))
    pt1 = (int(kp1[match[1], 0] + w0), int(kp1[match[1], 1]))
    if match_labels is None:
      color = tuple(rng.integers(0, 255, size=3).tolist())
    else:
      if match_labels[match[0], match[1]]:
        color = (0, 255, 0)
      else:
        color = (255, 0, 0)
    cv2.line(viz, pt0, pt1, color, line_width)

  # Optionally, add circles to output image to represent each keypoint.
  if show_keypoints:
    for i in range(np.shape(kp0)[0]):
      kp = kp0[i, :]
      if highlight_unmatched and has_unmatched_dustbins and match_matrix[i, -1]:
        cv2.circle(
            viz,
            tuple(kp.astype(np.int32).tolist()),
            circle_radius,
            (255, 0, 0),
            circle_thickness,
        )
      else:
        cv2.circle(
            viz,
            tuple(kp.astype(np.int32).tolist()),
            circle_radius,
            (0, 0, 255),
            circle_thickness,
        )
    for j in range(np.shape(kp1)[0]):
      kp = kp1[j, :]
      kp[0] += w0
      if highlight_unmatched and has_unmatched_dustbins and match_matrix[-1, j]:
        cv2.circle(
            viz,
            tuple(kp.astype(np.int32).tolist()),
            circle_radius,
            (255, 0, 0),
            circle_thickness,
        )
      else:
        cv2.circle(
            viz,
            tuple(kp.astype(np.int32).tolist()),
            circle_radius,
            (0, 0, 255),
            circle_thickness,
        )
  if title is not None:
    viz = cv2.putText(
        viz,
        title,
        (5, 30),
        cv2.FONT_HERSHEY_SIMPLEX,
        1,
        (0, 0, 255),
        2,
        cv2.LINE_AA,
    )
  return viz