Recognition with KNN is not really good

Hi Insightface team, thank you for sharing the code. I have tried to develop a recoginition demo using KNN. But I found the results is not good as I have done with face_recoginition of Dlib. The distance between faces and those in the model trained by archface is alway higher than face_recognition lib (mean that less confidence). Could you pls help me to review if I did smt wrong ? Here is the code I use to train :

`
X = []
y = []
parser = argparse.ArgumentParser(description='face model test')
parser.add_argument('--image-size', default='112,112', help='')
parser.add_argument('--model', default='./deploy/model-r100-ii/model,0', help='path to load model.')
parser.add_argument('--ga-model', default='./deploy/gamodel-r50/model,0', help='path to load model.')
parser.add_argument('--gpu', default=0, type=int, help='gpu id')
parser.add_argument('--det', default=0, type=int,
                    help='mtcnn option, 1 means using R+O, 0 means detect from begining')
parser.add_argument('--flip', default=0, type=int, help='whether do lr flip aug')
parser.add_argument('--threshold', default=1.24, type=float, help='ver dist threshold')
args = parser.parse_args()
model = face_model.FaceModel(args)
today = date.today().strftime("%d%m%Y")

# Loop through each person in the training set
i = 0;
for class_dir in os.listdir(train_dir):
    i += 1
    print(str(i) + '------------------' + class_dir)
    if not os.path.isdir(os.path.join(train_dir, class_dir)):
        continue

    # Loop through each training image for the current person
    for img_path in image_files_in_folder(os.path.join(train_dir, class_dir)):
        img = cv2.imread(img_path)
        img = model.get_input(img)
        print('Running inference... %s', img_path)
        if  img is None or len(img) == 0 :
             print('Cannot calculate ', img_path)
             continue
         start = time.time()
         _face_description = model.get_feature(img)
         print("Face extract took {} seconds.".format(time.time() - start))
         X.append(_face_description)
         y.append(class_dir)
clf = neighbors.KNeighborsClassifier(n_neighbors=n_neighbors, algorithm=knn_algo, weights='distance')
clf.fit(X, y)
return clf

` and the code to recognize is following:

  ```

   video_capture = cv2.VideoCapture(video_path + filename)
   ret, frame = video_capture.read()
   img = self.face_describer.get_input(frame)
   _face_description = self.face_describer.get_feature(img)
   _face_description = np.expand_dims(_face_description, axis=0)
   # Use the KNN model to find the best matches for the test face
   closest_distances = clf.kneighbors(_face_description, n_neighbors=5)
   are_matches = [closest_distances[0][i][0] <= distance_threshold for i in range(len(closest_distances[0]))]
   # Predict classes and remove classifications that aren't within the threshold
   return [(closest_distances[0][0][0], pred, loc) if rec else (closest_distances[0][0][0], "unknown", loc) for pred, loc, rec in zip(clf.predict(_face_description), X_face_locations, are_matches)]

Thanks. 

[tranvanhoa533]

If you use l2, you should normalize embedding first and threshold for distance is usually high ( about 1.0). If you use cosine, the threshold is smaller (about 0.5)

Vào Th 6, 19 thg 7, 2019 vào lúc 08:24 khanhnt [e-mail] đã viết:

which type of distance did you use ?

I used the defaul distance of sklearn’s KNN, that mean euclidean distance

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/deepinsight/insightface/issues/775?email_source=notifications&email_token=ADLMVXFS2JTFYKAP74D25ODQAEJU3A5CNFSM4H6MJG32YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD2KJCVA#issuecomment-513053012, or mute the thread https://github.com/notifications/unsubscribe-auth/ADLMVXENJ2YBRPZLEOWZ5BTQAEJU3ANCNFSM4H6MJG3Q .

[tomriddle54]

@khanhnt did you solve this using cosine as @tranvanhoa533 said if yes can you share your feedback on it. @tranvanhoa533 are you saying about cosine distance?