Update:

2020-03-28 Convert mobilenet0.25_Final.pth to caffemodel

Test

Compile my_caffe
Add PATH_of_YOUR_CAFFE_PYTHON and ../ to PYTHONPATH in toCaffe/detect.py：

from __future__ import print_function
import sys
sys.path.insert(0, "PATH_of_YOUR_CAFFE_PYTHON")
sys.path.insert(1, "../")
import caffe

Run detect.py

cd toCaffe
python detect.py

......
I0328 21:35:30.897866 15789 net.cpp:202] BatchNorm2d_7_BN does not need backward computation.
I0328 21:35:30.897871 15789 net.cpp:202] Conv2d_6 does not need backward computation.
I0328 21:35:30.897883 15789 net.cpp:202] LeakyReLU_5 does not need backward computation.
I0328 21:35:30.897891 15789 net.cpp:202] BatchNorm2d_4_Scale does not need backward computation.
I0328 21:35:30.897897 15789 net.cpp:202] BatchNorm2d_4_BN does not need backward computation.
I0328 21:35:30.897903 15789 net.cpp:202] Conv2d_3 does not need backward computation.
I0328 21:35:30.897910 15789 net.cpp:202] LeakyReLU_2 does not need backward computation.
I0328 21:35:30.897917 15789 net.cpp:202] BatchNorm2d_1_Scale does not need backward computation.
I0328 21:35:30.897923 15789 net.cpp:202] BatchNorm2d_1_BN does not need backward computation.
I0328 21:35:30.897930 15789 net.cpp:202] Conv2d_0 does not need backward computation.
I0328 21:35:30.897936 15789 net.cpp:202] data does not need backward computation.
I0328 21:35:30.897945 15789 net.cpp:244] This network produces output BboxHead_Concat
I0328 21:35:30.897953 15789 net.cpp:244] This network produces output ClassHead_Softmax
I0328 21:35:30.897960 15789 net.cpp:244] This network produces output LandmarkHead_Concat
I0328 21:35:30.898085 15789 net.cpp:257] Network initialization done.
Finished loading model!
net forward time: 1.0882
[650, 101, 704, 169, 0, 667, 127, 691, 128, 679, 141, 668, 151, 688, 151]
[661, 503, 714, 569, 0, 675, 528, 700, 526, 689, 538, 679, 551, 699, 549]
[596, 214, 655, 295, 0, 612, 244, 639, 242, 626, 255, 616, 273, 638, 271]
[143, 323, 202, 395, 0, 163, 345, 190, 344, 179, 361, 165, 375, 187, 375]
[520, 347, 576, 412, 0, 537, 366, 563, 367, 551, 379, 537, 392, 559, 393]
[100, 449, 148, 510, 0, 114, 471, 136, 470, 127, 484, 118, 494, 135, 494]
[575, 543, 634, 610, 0, 602, 562, 625, 565, 618, 577, 603, 590, 620, 593]
[697, 390, 747, 456, 0, 709, 412, 733, 410, 721, 420, 711, 436, 731, 434]
[760, 247, 812, 305, 0, 769, 269, 793, 269, 778, 284, 773, 291, 791, 291]
[110, 249, 167, 320, 0, 126, 272, 153, 272, 140, 284, 129, 301, 150, 300]
[375, 90, 428, 150, 0, 388, 113, 412, 110, 403, 122, 395, 136, 413, 133]
[227, 10, 276, 65, 0, 241, 27, 263, 26, 254, 37, 245, 49, 263, 48]
[835, 147, 890, 224, 0, 850, 175, 875, 175, 862, 190, 853, 204, 872, 204]
[741, 3, 789, 51, 0, 749, 18, 770, 15, 759, 27, 755, 38, 772, 36]
[303, 386, 367, 465, 0, 324, 414, 352, 416, 338, 430, 325, 444, 345, 446]
[74, 44, 116, 91, 0, 85, 57, 105, 58, 95, 66, 87, 76, 103, 77]
[379, 218, 447, 287, 0, 399, 238, 426, 238, 412, 248, 401, 264, 422, 265]
[421, 1, 461, 44, 0, 436, 10, 454, 8, 449, 16, 440, 29, 455, 28]
[127, 94, 165, 146, 0, 138, 115, 155, 116, 146, 126, 137, 134, 151, 135]
[923, 219, 980, 293, 0, 950, 239, 967, 244, 959, 256, 946, 270, 958, 275]
[793, 116, 837, 173, 0, 809, 133, 828, 135, 820, 144, 807, 156, 823, 157]
[294, 107, 334, 161, 0, 303, 128, 318, 127, 308, 137, 304, 148, 316, 147]
[60, 163, 112, 224, 0, 75, 181, 99, 180, 90, 188, 79, 204, 98, 203]
[951, 58, 979, 95, 0, 956, 71, 962, 72, 955, 80, 959, 87, 963, 87]
[990, -2, 1028, 39, 0, 1006, 3, 1025, 3, 1020, 14, 1008, 25, 1023, 25]
[800, 51, 840, 101, 0, 806, 68, 818, 68, 807, 78, 809, 88, 818, 88]

Below is origin result:

Loading pretrained model from ./weights/mobilenet0.25_Final.pth
remove prefix 'module.'
Missing keys:0
Unused checkpoint keys:0
Used keys:300
Finished loading model!
classifications: torch.Size([1, 26240, 2])
net forward time: 0.2385
[650, 101, 704, 169, 0, 667, 127, 691, 128, 679, 141, 668, 151, 688, 151]
[661, 503, 714, 569, 0, 675, 528, 700, 526, 689, 538, 679, 551, 699, 549]
[596, 214, 655, 295, 0, 612, 244, 639, 242, 626, 255, 616, 273, 638, 271]
[143, 323, 202, 395, 0, 163, 345, 190, 344, 179, 361, 165, 375, 187, 375]
[520, 347, 576, 412, 0, 537, 366, 563, 367, 551, 379, 537, 392, 559, 393]
[100, 449, 148, 510, 0, 114, 471, 136, 470, 127, 484, 118, 494, 135, 494]
[575, 543, 634, 610, 0, 602, 562, 625, 565, 618, 577, 603, 590, 620, 593]
[697, 390, 747, 456, 0, 709, 412, 733, 410, 721, 420, 711, 436, 731, 434]
[760, 247, 812, 305, 0, 769, 269, 793, 269, 778, 284, 773, 291, 791, 291]
[110, 249, 167, 320, 0, 126, 272, 153, 272, 140, 284, 129, 301, 150, 300]
[375, 90, 428, 150, 0, 388, 113, 412, 110, 403, 122, 395, 136, 413, 133]
[227, 10, 276, 65, 0, 241, 27, 263, 26, 254, 37, 245, 49, 263, 48]
[835, 147, 890, 224, 0, 850, 175, 875, 175, 862, 190, 853, 204, 872, 204]
[741, 3, 789, 51, 0, 749, 18, 770, 15, 759, 27, 755, 38, 772, 36]
[303, 386, 367, 465, 0, 324, 414, 352, 416, 338, 430, 325, 444, 345, 446]
[74, 44, 116, 91, 0, 85, 57, 105, 58, 95, 66, 87, 76, 103, 77]
[379, 218, 447, 287, 0, 399, 238, 426, 238, 412, 248, 401, 264, 422, 265]
[421, 1, 461, 44, 0, 436, 10, 454, 8, 449, 16, 440, 29, 455, 28]
[127, 94, 165, 146, 0, 138, 115, 155, 116, 146, 126, 137, 134, 151, 135]
[923, 219, 980, 293, 0, 950, 239, 967, 244, 959, 256, 946, 270, 958, 275]
[793, 116, 837, 173, 0, 809, 133, 828, 135, 820, 144, 807, 156, 823, 157]
[294, 107, 334, 161, 0, 303, 128, 318, 127, 308, 137, 304, 148, 316, 147]
[60, 163, 112, 224, 0, 75, 181, 99, 180, 90, 188, 79, 204, 98, 203]
[951, 58, 979, 95, 0, 956, 71, 962, 72, 955, 80, 959, 87, 963, 87]
[990, -2, 1028, 39, 0, 1006, 3, 1025, 3, 1020, 14, 1008, 25, 1023, 25]
[800, 51, 840, 101, 0, 806, 68, 818, 68, 807, 78, 809, 88, 818, 88]

RetinaFace in PyTorch

A PyTorch implementation of RetinaFace: Single-stage Dense Face Localisation in the Wild. Model size only 1.7M, when Retinaface use mobilenet0.25 as backbone net. We also provide resnet50 as backbone net to get better result. The official code in Mxnet can be found here.

Mobile or Edge device deploy

We also provide a set of Face Detector for edge device in here from python training to C++ inference.

WiderFace Val Performance in single scale When using Resnet50 as backbone net.

Style	easy	medium	hard
Pytorch (same parameter with Mxnet)	94.82 %	93.84%	89.60%
Pytorch (original image scale)	95.48%	94.04%	84.43%
Mxnet	94.86%	93.87%	88.33%
Mxnet(original image scale)	94.97%	93.89%	82.27%

WiderFace Val Performance in single scale When using Mobilenet0.25 as backbone net.

Style	easy	medium	hard
Pytorch (same parameter with Mxnet)	88.67%	87.09%	80.99%
Pytorch (original image scale)	90.70%	88.16%	73.82%
Mxnet	88.72%	86.97%	79.19%
Mxnet(original image scale)	89.58%	87.11%	69.12%

FDDB Performance.

FDDB(pytorch)	performance
Mobilenet0.25	98.64%
Resnet50	99.22%

Installation
Training
Evaluation
References

Installation

Clone and install

git clone https://github.com/biubug6/Pytorch_Retinaface.git
Pytorch version 1.1.0+ and torchvision 0.3.0+ are needed.
Codes are based on Python 3

Data

Download the WIDERFACE dataset.
Download annotations (face bounding boxes & five facial landmarks) from baidu cloud or dropbox
Organise the dataset directory as follows:

  ./data/widerface/
    train/
      images/
      label.txt
    val/
      images/
      wider_val.txt

ps: wider_val.txt only include val file names but not label information.

Data1

We also provide the organized dataset we used as in the above directory structure.

Link: from google cloud or baidu cloud Password: ruck

Training

We provide restnet50 and mobilenet0.25 as backbone network to train model. We trained Mobilenet0.25 on imagenet dataset and get 46.58% in top 1. If you do not wish to train the model, we also provide trained model. Pretrain model and trained model are put in google cloud and baidu cloud Password: fstq . The model could be put as follows:

  ./weights/
      mobilenet0.25_Final.pth
      mobilenetV1X0.25_pretrain.tar
      Resnet50_Final.pth

Before training, you can check network configuration (e.g. batch_size, min_sizes and steps etc..) in data/config.py and train.py.
Train the model using WIDER FACE:

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --network resnet50 or
CUDA_VISIBLE_DEVICES=0 python train.py --network mobile0.25

Evaluation

Evaluation widerface val

Generate txt file

python test_widerface.py --trained_model weight_file --network mobile0.25 or resnet50

Evaluate txt results. Demo come from Here

cd ./widerface_evaluate
python setup.py build_ext --inplace
python evaluation.py

You can also use widerface official Matlab evaluate demo in Here

Evaluation FDDB

Download the images FDDB to:

./data/FDDB/images/

Evaluate the trained model using:

python test_fddb.py --trained_model weight_file --network mobile0.25 or resnet50

Download eval_tool to evaluate the performance.

References

@inproceedings{deng2019retinaface,
title={RetinaFace: Single-stage Dense Face Localisation in the Wild},
author={Deng, Jiankang and Guo, Jia and Yuxiang, Zhou and Jinke Yu and Irene Kotsia and Zafeiriou, Stefanos},
booktitle={arxiv},
year={2019}

jnulzl/Pytorch_Retinaface_To_Caffe