sierkinhane/mtcnn-pytorch

A face detection algorithm

mtcnn-pytorch

Descriptions in chinese

https://blog.csdn.net/Sierkinhane/article/details/83308658

results:

Test an image

• run > python mtcnn_test.py

Training data prepraring

• download WIDER FACE (passcode:lsl3) face detection data then store it into ./data_set/face_detection
  • run > python ./anno_store/tool/format/transform.py change .mat(wider_face_train.mat) into .txt(anno_train.txt)
• download CNN_FacePoint face detection and landmark data then store it into ./data_set/face_landmark

Training

• preparing data for P-Net

  • run > python mtcnn/data_preprocessing/gen_Pnet_train_data.py
  • run > python mtcnn/data_preprocessing/assemble_pnet_imglist.py

• train P-Net

  • run > python mtcnn/train_net/train_p_net.py

• preparing data for R-Net

  • run > python mtcnn/data_preprocessing/gen_Rnet_train_data.py (maybe you should change the pnet model path)
  • run > python mtcnn/data_preprocessing/assemble_rnet_imglist.py

• train R-Net

  • run > python mtcnn/train_net/train_r_net.py

• preparing data for O-Net

  • run > python mtcnn/data_preprocessing/gen_Onet_train_data.py
  • run > python mtcnn/data_preprocessing/gen_landmark_48.py
  • run > python mtcnn/data_preprocessing/assemble_onet_imglist.py

• train O-Net

  • run > python mtcnn/train_net/train_o_net.py

Citation

DFace

mtcnn-pytorch's Issues

some errors find in the code

1. wider_load generate a error anno.txt for image path
2. wider_load generate a error anno.txt for image name

result on fddb

hi，have u evaluated the model on FDDB dataset，what is the result

The normalize

The normalize coordinate of face detector and landmark somehow is not correct.

do you have a pretrained models (trained with widerface) to use it right away ?

Can you upload pre-trained models (trained with wider-faces) to use it for detecting right away ?

gen_rnet_data

Hi, I've been trying to train this model from scratch on a local dataset, but I've been encountering a CUDA memory issue when i run gen_Rnet_train_data.py
I'm using a 1050ti , 4gb GPU with 16 gb RAM.

Traceback (most recent call last): File "mtcnn/data_preprocess/gen_Pnet_train_data.py", line 94, in <module> delta_x = np.random.randint(max(-size, -x1), w) File "mtrand.pyx", line 630, in numpy.random.mtrand.RandomState.randint File "bounded_integers.pyx", line 1228, in numpy.random.bounded_integers._rand_int64 ValueError: low >= high

Hi, when I generate the Pnet train data, some errors happend:

12880 pics in total
/media/gaoya/disk/Applications/pytorch/目标检测/人脸检测/dataset/WIDER_FACE/WIDER_train/images/0--Parade/0_Parade_marchingband_1_849.jpg
/media/gaoya/disk/Applications/pytorch/目标检测/人脸检测/dataset/WIDER_FACE/WIDER_train/images/0--Parade/0_Parade_Parade_0_904.jpg
Traceback (most recent call last):
  File "mtcnn/data_preprocess/gen_Pnet_train_data.py", line 94, in <module>
    delta_x = np.random.randint(max(-size, -x1), w)
  File "mtrand.pyx", line 630, in numpy.random.mtrand.RandomState.randint
  File "bounded_integers.pyx", line 1228, in numpy.random.bounded_integers._rand_int64
ValueError: low >= high

How can I solve it?

anno_train.txt has nothing!

landmark data not generated for pnet and rnet

@Sierkinhane why are we not generating landmark data for pnet and rnet ?

找不到‘’./anno_store/anno_train_test.txt‘’文件

请问在
https://github.com/Sierkinhane/mtcnn-pytorch/blob/master/mtcnn/data_preprocess/gen_Rnet_train_data.py#L23，
怎样得到./anno_store/anno_train_test.txt

WIDER FACE data set link is not working

https://pan.baidu.com/s/1sJTO7TcQ2576RUqR_IIhbQ

人脸框太大了

人脸框太大了`````

rnet时生成的rnet_epoch_1.pt-rnet_epoch_10.pt人脸一个也检测不到

你好，问下按照你的步骤在wider 数据集上进行训练，在选用pnet_epoch_10.pt 去训练rnet时生成的rnet_epoch_1.pt-rnet_epoch_10.pt人脸一个也检测不到。

“./” 寻找root路劲无效

在第一步，生成pnet数据，所有的路劲都不正确，似乎因为'./'不能被识别

Hard Sample Mining

Thx for your implementation, what a piece of beautiful work!
However, I notice that Hard Sample Mining is not implemented in your code,
can u share ur idea about HSM?
Thank u!!!

如何返回坐标

关于 gen_rnt_train_data中annotation_file

你好我想问下annotation_file 是整个数据集的txt文件路径吗，是否和gen_pnt_train_data 中的 anno_file 是同一个文件

Onet trained model not working

I have trained onet by following exact same steps on the same dataset.
But when I have replaced it with onet from the demo example. it is not working.

在训练rnet的时候loss会爆炸 有人遇到这个问题吗

how to generate anno_train_test.txt

when run > python mtcnn/data_preprocessing/gen_Rnet_train_data.py

Path does not exist: ./anno_store/anno_train_test.txt