cunjian / pytorch_face_landmark

Fast and accurate face landmark detection library using PyTorch; Support 68-point semi-frontal and 39-point profile landmark detection; Support both coordinate-based and heatmap-based inference; Up to 100 FPS landmark inference speed with SOTA face detector on CPU.
866 stars 157 forks source link

detection or prediction #9

Closed BehnazDibayee closed 4 years ago

BehnazDibayee commented 4 years ago

Hi Thanks for sharing your work. I test "test_camera_mtcnn_onnx.py" with my webcam and block part of my face with a mask (my mouth and nose). I expect the code to reduce keypoints and only show some of them which were related to lower part of my face. But it didn't. It still showed all keypoints inaccurately and seems to predict them instead of detecting. Is that a detection or a prediction?

cunjian commented 4 years ago

I assume you use the 56*56 MobileNetV2 model. This model has not been trained with occlusion data, nor does it include any occlusion indices. Basically, it will output all 68 landmarks. There is not much difference between detection and prediction. These points are all estimated. You can try the PFLD model. It is better against occlusion. Maybe in the future, I can add the artificially generated mask onto the face, this should help address the problem you mention.