This is a course project i have created using pytorch ,using all the skills i learned from freecodecamp and Jovian.
Run the Detector_MTCNN.py file. At present video is taken from the webcam(live) if you want to feed in a pre-recorded video give the path of the file instead of 0 in line 28 cv.VideoCapture(0). If the video is too big and potentially freeze the computer uncomment line 57 #frame = resize(frame, height, width) this will resize it.
Make sure to download the state dict to get the predictions right
state dict- https://drive.google.com/drive/folders/1oRBDw_HmqCaQ2jnT4aSZHyYVBi4ELhSt?usp=sharing,
dataset - https://drive.google.com/drive/folders/1LEKdePxk854r0kT542g42loM1z1UkL4g?usp=sharing
I tried both the models with different video's, ResNet9 and ResNet15 performed well. I noticed that there are some video ResNet9 performed well but ResNet15 did not and vice-versa.
Try both the models and see whats best.
The model is trained on certain type of mask so it may not perform well on other kinds of mask.
- Facenet PyTorch
- Open CV
- PyTorch
- Numpy
- Matplotlib
- Guide to MTCNN in facenet-pytorch - https://www.kaggle.com/timesler/guide-to-mtcnn-in-facenet-pytorch
- Facenet implementation in a video - https://github.com/timesler/facenet-pytorch/blob/master/examples/face_tracking.ipynb
,
