Improved MTCNN TF Training

This repository contains an improved version of MTCNN training using TensorFlow. It is based on AITTSMD/MTCNN-Tensorflow, which reproduces the original MTCNN paper. The main enhancements are:

• Simpler and clearer code structure
• Easy to execute
• Compatible with TensorFlow 2.11 or lower
• Support for CUDA GPU acceleration
• Python 3.10 compatible

All you need is the training dataset from Wider Face Training Dataset and CNN Facepoint Training Dataset, then put it in the dataset folder.

Key Points :

• gen_hard_bbox_pnet.py is used to generate 12x12 training images for PNet
• gen_hard_bbox_rnet_onet.py is used to generate 24x24 training images for RNet, and 48x48 training images for ONet
• gen_landmark_aug.py is used to generate 12x12, 24x24, and 48x48 facial landmark images for all the networks
• training.py is used to train PNet, RNet, and ONet
• gen_tfrecords.py is used to save training data in tfrecord format for easy training

Training Steps：

• PNet
  1. Run python gen_hard_bbox_pnet.py
  2. Run python gen_landmark_aug.py --stage pnet
  3. Run python gen_tfrecords.py --stage pnet
  4. Run python train.py --stage pnet
• RNet 5. Run python gen_hard_bbox_rnet_onet.py --stage rnet 5. Run python gen_landmark_aug.py --stage rnet 5. Run python gen_tfrecords.py --stage rnet 5. Run python train.py --stage rnet
• ONet 9. Run python gen_hard_bbox_rnet_onet.py --stage onet 9. Run python gen_landmark_aug.py --stage onet 9. Run python gen_tfrecords.py --stage onet 9. Run python train.py --stage onet

And you're done! Enjoy the results! :)i

#FDDB Result (Face Detection Dataset and Benchmark):

True positive rate - 0.93 False positive - 422

#Wider Face Detection Benchmark:

wider evaluation easy - 0.803 medium - 0.775 hard - 0.508

**EASY SET**

**MEDIUM SET**

**HARD SET**