bobliu20/mtcnn_tf

MTCNN implement by tensorflow. Easy to training and testing.

Description

MTCNN: Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks
Implement training and testing by tensorflow.

Dependencies

• Tensorflow v1.0.0 or higher
• TF-Slim
• Python 2.7
• Ubuntu 14.04 or CentOS 7.2 or higher
• Cuda 8.0

Prepare Data and Start Training

1. WIDER face dataset: Download WIDER_train.zip from here. You can only download Wider Face Training Images. Unzip it and move it to dataset folder.
2. Landmark dataset: Download train.zip from here. You can only download training set. Unzip it and move it to dataset folder.
3. Run ./clearAll.sh to clear all tmp file.
4. Run ./runAll.sh to finish all(Include prepare data and training). Please check this shell script to get more info.

Testing and predict

1. Copy your image file to testing/images
2. Run python testing/test_images.py --stage=onet. Anyway you can specify stage to pnet or rnet to check your model.
3. The result will output in testing/results_onet

Results

License

MIT LICENSE

References

1. Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, Yu Qiao , " Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks," IEEE Signal Processing Letter
2. MTCNN-Tensorflow

mtcnn_tf's Issues

Training my own dataset

if I have to train my own dataset, How do I do that? how the dataset has to be?

assert(img is not None)

hello Bob, when I train the face data,there is an error: assert(img is not None). but i can show the image in the code, i don't know where is the bug? can u help me??? thanks.

Error in gen_hard_bbox_rnet_onet.py

Hello Bob, thank you for such nice code. your repo is much cleaner than other versions. Hovewer, I found one tiny error in gen_hard_bbox_rnet_onet.py in line 191:
minFace = 24

For ONet stage minFace valuer should be 48 I think. If use 24 it fails to generate tfrecords, cause a lot of jpg files don't exist in data/onet folder. training also become meaningless with this setting.

I've changed minFace value to 48 and this solved this problem.

May be it is a mistake in file gen_hard_bbox_pnet.py

2. NEG: random to crop sample image in bbox inside

In the line 84， iou should be Iou

      83：      Iou = IoU(crop_box, boxes)
      84：      if np.max(iou) >= 0.3:
                          continue

questions about landmark location

thank you very much for the project. its extremely helpful : )
however, it seems the location is slightly inaccurate. do you think a larger input with a more complex net would helpful? and when i was to detect more landmarks, is it likely to retain the accuracy?

关于P-Net检测的疑问

请教P-Net检测：输入经resize后的图片检测成功后，需将得到的box坐标还原至原图上，这个操作中乘上stride是什末意思 ？
代码如下
boundingbox = np.vstack([np.round((self.stride * t_index[1]) / scale),
np.round((self.stride * t_index[0]) / scale),
np.round((self.stride * t_index[1] + cellsize) / scale),
np.round((self.stride * t_index[0] + cellsize) / scale),
score,
reg])

Request for code review.

Please review MTCNN TensorFlow code at https://github.com/look4pritam/TensorFlowMTCNN on CelebA branch.