mtcnn How could I use your `train.py` in the `example` file

I want to train some pictures using mtcnn algorithm. And how could I use train.py, could you tell me some steps to use it?

Dec 28 '16 08:12 mycastiel

@mycastiel train.py cannot be used directly. You shall use example/train_P/R/O_net.py to train the three networks, and use prepare_data/gen_hard_example.py to collect training data for next stage.

Dec 28 '16 11:12 Seanlinx

Thx, it really helps me something. However, these files' dataset_path has the default value of data/***, where should I point it to and where should the diretory located? Thank you

Dec 29 '16 02:12 mycastiel

You can have a look at README.md where I explained the data folder structure.

Dec 29 '16 06:12 Seanlinx

Thx. I almost perfectly run it now. But, I find that I don't know what should be put in anno.txt, for I just generated a empty file named anno.txt. Where could I find something similar of it or could generate it.

Dec 29 '16 08:12 mycastiel

#3

Dec 29 '16 09:12 Seanlinx

But the anno.txt doesn't use the dataset FDDB you provide, and I can't find the dataset in github or Baidu

Dec 29 '16 09:12 mycastiel

And the anno file FDDB provides doesn't fit your codes. For instance some lines only contains one String, but the code needs to read more than three strings

Dec 29 '16 09:12 mycastiel

The training set is WIDER FACE and this code is for wider face not FDDB. If you're going to train FDDB, you should generate anno.txt from FDDB anno files, transform the ellipse annotations into rectangles, and write into format: image_name x1 y1 x2 y2 ... an example would be:
2002/08/09/big/img_242 238.081330833 15.9869071934 394.707127167 257.596796807 -43.3222110851 107.890378739 83.3887630851 304.440661261

Dec 29 '16 10:12 Seanlinx

Hi, Seanlinux. I really appreciate your codes. I am new in mxnet, so I get stuck at the first step, datasets... I just want to train P/R/O net on the dataset of wider_face, I put it at /data/wider. but I am not really understand what do "data/mtcnn" and "data/wider" represent? In the "train_P_net.py", the default dataset_path is "data/mtcnn", while in the "gen_hard_example.py", the default dataset_path is "data/wider", could you please explain it and tall me the right place I should put the wider face dataset. Thank you ! @Seanlinx

Dec 01 '17 07:12 anna0709

@anna0709 Place the wider face dataset in "data/wider". "data/mtcnn" stores training images cropped from wider face dataset. You can refer to readme for more details

Dec 01 '17 08:12 Seanlinx