YJingyu / tf_JDAP

Joint face detection and face landmark alignment and head pose evaluation using CNN and cascade structure.

JDAP(Joint Detection and Ailgnment and Head Pose)

A face detection algorithm joint multi-task using cascade structure in CNN. This repository containtraining code and testing code using Tensorflow architecture.

Results

Requirement

Quick overview of requirements: - Linux(ubuntu 16.04 my test) - tensorflow 1.0.1(or more high) - python 2.7 - opencv-python - easydict

Preparation Data

- WIDER-FACE
- 300W-LP + Menpo-3D or (CelebA + AFLW)

Data Label Design

Classification task:

Positive, Part, Negative sample label : image_name class_id bounding_box_regressor class_id: +1(Positive) -1(Part) 0(Negative) Negative bbox_regressor: 0 0 0 0(anything, but must keep 4 number) bounding_box_regressor: x1 y1 x2 y2(relative ground truth)

Auxiliary task:

Landmark sample label: image_name class_id landmark_regressor Pose sample label: image_name class_id pose_regressor class_id: -2(Landmark) -3(Pose)

Training PNet

1. Modfiy data_root_dir and save_dir_root of "prepare_data/gen_pnet_train_data.py" and "prepare_data/gen_pnet_val_data.py" .
2. Select suitable parameters.

• Main options:
  • IoU thresh
  • how many negative samples per image
  • pos_aug_ratio(without augment model performance is good. T^T)

3. ./scripts/make_pnet_train_val_data.sh

4. Modfiy "prepare_data/gen_imglist.py" **save_data_dir **and netSize
5. According to self task design label in "prepare_data/multithread_create_tfrecords.py" _set_single_example
6. Modify "scripts/make_tfrecords.sh" and run
7. Adjustment "scripts/train_cls.sh" and run

Training RNet

1. Modfiy "prepare_data/gen_hard_sample.py" set yourself fold root
2. stage = 1 get PNet detect result and save it as pickle
3. stage = 2 crop and save image patch
4. Like PNet generate tfrecords
5. Adjustment "scripts/train_cls.sh" and run

Training ONet

1. Modfiy "prepare_data/gen_hard_sample.py" set yourself fold root, get classify samples

Demo

Select model file and suitable hyper parameters

python ./demo/mtcnn.py

FAQ

1. lr = 0.01, lr_decay_scale = 0.1 and epoch [7 ,13] make lr decay

2. Small batch size(BS)

   • Small BS can get higher recall than large BS in FDDB.

3. l2 regularizer is small

   • Change normal 5e^-5 to 1e^-5, network less limit.

4. Add part samples in trianing stage

   • Help to bounding box regression and indirect promote face classification.

5. Less channel

   • There's negligible advance using more channel.

6. Optimizer

   • Momentum optimizer in 0.9 momentum.

7. OHEM achieve, ohem ratio is 0.7

   • BP top 70% of loss. The loss except part samples.

8. Focal loss VS. SoftmaxWithLoss

   • SF is more less false positive and higher recall than FL.

9. ERC(Early recject classifier) and DR Layer

   • New strategy in [2], but increase more parameters.

Future work

1. PNet use conv1-s2 replace max pooling and relu6 replace prelu. But relu6 and reduce max pooling method lead to more false positive ratio and lower recall.

References

1. K. Zhang, Z. Zhang, Z. Li and Y. Qiao. Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks. Signal Processing Letters, 23(10):1499–1503, 2016.
2. K. Zhang, Z. Zhang, Z. Li and Y. Qiao. Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks. Signal Processing Letters, 23(10):1499–1503, 2016.
3. V. Jain and E. Learned-Miller. FDDB: A benchmark for face detection in unconstrained settings. In Technical Report UMCS-2010-009, 2010.
4. S. Yang, P. Luo, C.-C. Loy, and X. Tang. Wider face: A face detection benchmark. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016.