senadkurtisi / MTCNN-pytorch

PyTorch implementation of Multi Task CNN paper for face detection

MTCNN-pytorch

PyTorch implementation of Multi Task CNN paper for face detection.

The algorithm

The algorithm performs 3 stage detection process. Each stage uses different CNN, therefore this detection system is represented as a cascade system of three CNNs.

Stages:

1. First CNN: P-Net (Proposal Network) produces initial guesses (candidates) for bounding boxes of faces, along with probabilities that those boxes contain a face. Some of the proposed candidates are rejected due to probability threshold and NMS. Bounding boxes are calibrated
2. Second CNN: R-Net (Refinement Network) refines former candidates in order to get a better estimate. Some of the proposed candidates are rejected due to probability threshold and NMS. Bounding boxes are calibrated
3. Third CNN: O-net (Output Network) outputs final estimates of bounding boxes and according facial landmark positions

CNNs are pretrained and used as such. This repo is an implementation of Zhang's work. The weights can be found in the pretrained weights folder.

Results

Instructions

1. Open Anaconda Prompt and navigate to the directory of this repo by using: cd PATH_TO_THIS_REPO
2. Execute conda env create -f environment.yml This will set up an environment with all necessary dependencies.
3. Activate previously created environment by executing: conda activate mtcnn-pytorch
4. Start the main script for face detection: python src/main.py --img_loc PATH_TO_IMAGE or just execute python src/main.py and use default locations for image. Default location is: imgs/girls.jpg

References

This repo is heavily inspired by these two repositories:

1. MX Net implementation of MTCNN
2. PyTorch implementation