Difference on accuracy between your results and my reproduction

[on-your-way]

Hi, I have tried to reproduce your results of LFFD v1 on the WIDER_FACE_val dataset. The training data I used is produced by your method:

1. Input the WIDER FACE train dataset (downloaded from the official website);
2. Generate the train_list.txt file by reformat_WIDERFACE.py;

3. Generate the .pkl file by pickle_provider.py And then, run configuration_10_560_25L_8scales_v1.py (specify the .pkl file generated just now). The only parameter I have modified is the param_num_thread_train_dataiter at the line 56 (change it from 4 to 10). After 1,400,000 iterations, I evaluated it on the WIDER_FACE_val dataset using standard tools released by the WIDER FACE benchmark. However, my results of accuracy are quite lower than yours, which are obtained by face_detection/saved_model/configuration_10_560_25L_8scales_v1/train_10_560_25L_8scales_v1_iter_1400000.params you provided in the project. Here are details:

        easy      medium   hard

   mine 0.885 0.853 0.681 yours 0.896 0.861 0.710 diff -1.2% -0.9% -4.1%

   I can not figure it out really. May you give me some help? thanks

[antoniojmabreu]

Sorry, I know my questio is offtopic, but can you help me run LFD on windows? I get an error using libturbojpeg.

(env) PS G:\OneDrive - Universidade da Beira Interior\EI\Mestrado\1 ano\2 semestre\IC\Projeto\LFD-A-Light-and-Fast-Detector-master\WIDERFACE_train> python.exe .\predict.py Traceback (most recent call last): File ".\predict.py", line 4, in from lfd.data_pipeline.augmentation import File "G:\OneDrive - Universidade da Beira Interior\EI\Mestrado\1 ano\2 semestre\IC\Projeto\LFD-A-Light-and-Fast-Detector-master\WIDERFACE_train\lfd\data_pipeline__init__.py", line 4, in from .data_loader import File "G:\OneDrive - Universidade da Beira Interior\EI\Mestrado\1 ano\2 semestre\IC\Projeto\LFD-A-Light-and-Fast-Detector-master\WIDERFACE_train\lfd\data_pipeline\data_loader__init.py", line 3, in from .data_loader import DataLoader File "G:\OneDrive - Universidade da Beira Interior\EI\Mestrado\1 ano\2 semestre\IC\Projeto\LFD-A-Light-and-Fast-Detector-master\WIDERFACE_train\lfd\data_pipeline\data_loader\data_loader.py", line 8, in from ..dataset import turbojpeg, reserved_keys File "G:\OneDrive - Universidade da Beira Interior\EI\Mestrado\1 ano\2 semestre\IC\Projeto\LFD-A-Light-and-Fast-Detector-master\WIDERFACE_train\lfd\data_pipeline\dataset__init.py", line 7, in from .utils import * File "G:\OneDrive - Universidade da Beira Interior\EI\Mestrado\1 ano\2 semestre\IC\Projeto\LFD-A-Light-and-Fast-Detector-master\WIDERFACE_train\lfd\data_pipeline\dataset\utils\init__.py", line 3, in from .turbojpeg import turbojpeg File "G:\OneDrive - Universidade da Beira Interior\EI\Mestrado\1 ano\2 semestre\IC\Projeto\LFD-A-Light-and-Fast-Detector-master\WIDERFACE_train\lfd\data_pipeline\dataset\utils\turbojpeg.py", line 456, in turbojpeg = TurboJPEG(lib_path=target_libturbojpeg_path) File "G:\OneDrive - Universidade da Beira Interior\EI\Mestrado\1 ano\2 semestre\IC\Projeto\LFD-A-Light-and-Fast-Detector-master\WIDERFACE_train\lfd\data_pipeline\dataset\utils\turbojpeg.py", line 153, in init self.find_turbojpeg() if lib_path is None else lib_path) File "C:\Users\anton\AppData\Local\Programs\Python\Python37\lib\ctypes__init__.py", line 442, in LoadLibrary return self._dlltype(name) File "C:\Users\anton\AppData\Local\Programs\Python\Python37\lib\ctypes__init.py", line 364, in init__ self._handle = _dlopen(self._name, mode) OSError: [WinError 193] %1 is not a valid Win32 application

I did build it using Visual Studio 16, but it isn't working.

Thanks in advance.

[on-your-way]

@antoniojmabreu I have encountered many errors on Windows just like you, so I turn to Linux, and then everything works well.

[antoniojmabreu]

Thank you!

I will switch to Linux then.

[gkrisp98]

I believe this drop in accuracy is normal