Add SFace visualization demo and example outputs (#231)

README.md

@@ -61,6 +61,10 @@ Some examples are listed below. You can find more in the directory of each model
 
 ![largest selfie](./models/face_detection_yunet/example_outputs/largest_selfie.jpg)
 
+### Face Recognition with [SFace](./models/face_recognition_sface/)
+
+![sface demo](./models/face_recognition_sface/example_outputs/demo.jpg)
+
 ### Facial Expression Recognition with [Progressive Teacher](./models/facial_expression_recognition/)
 
 ![fer demo](./models/facial_expression_recognition/example_outputs/selfie.jpg)

models/face_recognition_sface/README.md

@@ -26,12 +26,18 @@ Run the following command to try the demo:
 
 ```shell
 # recognize on images
-python demo.py --input1 /path/to/image1 --input2 /path/to/image2
+python demo.py --target /path/to/image1 --query /path/to/image2
 
 # get help regarding various parameters
 python demo.py --help
 ```
 
+### Example outputs
+
+![sface demo](./example_outputs/demo.jpg)
+
+Note: Left part of the image is the target identity, the right part is the query. Green boxes are the same identity, red boxes are different identities compared to the left.
+
 ## License
 
 All files in this directory are licensed under [Apache 2.0 License](./LICENSE).

models/face_recognition_sface/demo.py

@@ -30,10 +30,10 @@ backend_target_pairs = [
 
 parser = argparse.ArgumentParser(
     description="SFace: Sigmoid-Constrained Hypersphere Loss for Robust Face Recognition (https://ieeexplore.ieee.org/document/9318547)")
-parser.add_argument('--input1', '-i1', type=str,
-                    help='Usage: Set path to the input image 1 (original face).')
-parser.add_argument('--input2', '-i2', type=str,
-                    help='Usage: Set path to the input image 2 (comparison face).')
+parser.add_argument('--target', '-t', type=str,
+                    help='Usage: Set path to the input image 1 (target face).')
+parser.add_argument('--query', '-q', type=str,
+                    help='Usage: Set path to the input image 2 (query).')
 parser.add_argument('--model', '-m', type=str, default='face_recognition_sface_2021dec.onnx',
                     help='Usage: Set model path, defaults to face_recognition_sface_2021dec.onnx.')
 parser.add_argument('--backend_target', '-bt', type=int, default=0,
@@ -46,8 +46,64 @@ parser.add_argument('--backend_target', '-bt', type=int, default=0,
                     '''.format(*[x for x in range(len(backend_target_pairs))]))
 parser.add_argument('--dis_type', type=int, choices=[0, 1], default=0,
                     help='Usage: Distance type. \'0\': cosine, \'1\': norm_l1. Defaults to \'0\'')
+parser.add_argument('--save', '-s', action='store_true',
+                    help='Usage: Specify to save file with results (i.e. bounding box, confidence level). Invalid in case of camera input.')
+parser.add_argument('--vis', '-v', action='store_true',
+                    help='Usage: Specify to open a new window to show results. Invalid in case of camera input.')
 args = parser.parse_args()
 
+def visualize(img1, faces1, img2, faces2, matches, scores, target_size=[512, 512]): # target_size: (h, w)
+    out1 = img1.copy()
+    out2 = img2.copy()
+    matched_box_color = (0, 255, 0)    # BGR
+    mismatched_box_color = (0, 0, 255) # BGR
+
+    # Resize to 256x256 with the same aspect ratio
+    padded_out1 = np.zeros((target_size[0], target_size[1], 3)).astype(np.uint8)
+    h1, w1, _ = out1.shape
+    ratio1 = min(target_size[0] / out1.shape[0], target_size[1] / out1.shape[1])
+    new_h1 = int(h1 * ratio1)
+    new_w1 = int(w1 * ratio1)
+    resized_out1 = cv.resize(out1, (new_w1, new_h1), interpolation=cv.INTER_LINEAR).astype(np.float32)
+    top = max(0, target_size[0] - new_h1) // 2
+    bottom = top + new_h1
+    left = max(0, target_size[1] - new_w1) // 2
+    right = left + new_w1
+    padded_out1[top : bottom, left : right] = resized_out1
+
+    # Draw bbox
+    bbox1 = faces1[0][:4] * ratio1
+    x, y, w, h = bbox1.astype(np.int32)
+    cv.rectangle(padded_out1, (x + left, y + top), (x + left + w, y + top + h), matched_box_color, 2)
+
+    # Resize to 256x256 with the same aspect ratio
+    padded_out2 = np.zeros((target_size[0], target_size[1], 3)).astype(np.uint8)
+    h2, w2, _ = out2.shape
+    ratio2 = min(target_size[0] / out2.shape[0], target_size[1] / out2.shape[1])
+    new_h2 = int(h2 * ratio2)
+    new_w2 = int(w2 * ratio2)
+    resized_out2 = cv.resize(out2, (new_w2, new_h2), interpolation=cv.INTER_LINEAR).astype(np.float32)
+    top = max(0, target_size[0] - new_h2) // 2
+    bottom = top + new_h2
+    left = max(0, target_size[1] - new_w2) // 2
+    right = left + new_w2
+    padded_out2[top : bottom, left : right] = resized_out2
+
+    # Draw bbox
+    assert faces2.shape[0] == len(matches), "number of faces2 needs to match matches"
+    assert len(matches) == len(scores), "number of matches needs to match number of scores"
+    for index, match in enumerate(matches):
+        bbox2 = faces2[index][:4] * ratio2
+        x, y, w, h = bbox2.astype(np.int32)
+        box_color = matched_box_color if match else mismatched_box_color
+        cv.rectangle(padded_out2, (x + left, y + top), (x + left + w, y + top + h), box_color, 2)
+
+        score = scores[index]
+        text_color = matched_box_color if match else mismatched_box_color
+        cv.putText(padded_out2, "{:.2f}".format(score), (x + left, y + top - 5), cv.FONT_HERSHEY_DUPLEX, 0.4, text_color)
+
+    return np.concatenate([padded_out1, padded_out2], axis=1)
+
 if __name__ == '__main__':
     backend_id = backend_target_pairs[args.backend_target][0]
     target_id = backend_target_pairs[args.backend_target][1]
@@ -65,17 +121,35 @@ if __name__ == '__main__':
                      backendId=backend_id,
                      targetId=target_id)
 
-    img1 = cv.imread(args.input1)
-    img2 = cv.imread(args.input2)
+    img1 = cv.imread(args.target)
+    img2 = cv.imread(args.query)
 
     # Detect faces
     detector.setInputSize([img1.shape[1], img1.shape[0]])
-    face1 = detector.infer(img1)
-    assert face1.shape[0] > 0, 'Cannot find a face in {}'.format(args.input1)
+    faces1 = detector.infer(img1)
+    assert faces1.shape[0] > 0, 'Cannot find a face in {}'.format(args.target)
     detector.setInputSize([img2.shape[1], img2.shape[0]])
-    face2 = detector.infer(img2)
-    assert face2.shape[0] > 0, 'Cannot find a face in {}'.format(args.input2)
+    faces2 = detector.infer(img2)
+    assert faces2.shape[0] > 0, 'Cannot find a face in {}'.format(args.query)
 
     # Match
-    result = recognizer.match(img1, face1[0][:-1], img2, face2[0][:-1])
-    print('Result: {}.'.format('same identity' if result else 'different identities'))
+    scores = []
+    matches = []
+    for face in faces2:
+        result = recognizer.match(img1, faces1[0][:-1], img2, face[:-1])
+        scores.append(result[0])
+        matches.append(result[1])
+
+    # Draw results
+    image = visualize(img1, faces1, img2, faces2, matches, scores)
+
+    # Save results if save is true
+    if args.save:
+        print('Resutls saved to result.jpg\n')
+        cv.imwrite('result.jpg', image)
+
+    # Visualize results in a new window
+    if args.vis:
+        cv.namedWindow("SFace Demo", cv.WINDOW_AUTOSIZE)
+        cv.imshow("SFace Demo", image)
+        cv.waitKey(0)

models/face_recognition_sface/sface.py

@@ -57,7 +57,7 @@ class SFace:
 
         if self._disType == 0: # COSINE
             cosine_score = self._model.match(feature1, feature2, self._disType)
-            return 1 if cosine_score >= self._threshold_cosine else 0
+            return cosine_score, 1 if cosine_score >= self._threshold_cosine else 0
         else: # NORM_L2
             norml2_distance = self._model.match(feature1, feature2, self._disType)
-            return 1 if norml2_distance <= self._threshold_norml2 else 0
+            return norml2_distance, 1 if norml2_distance <= self._threshold_norml2 else 0