DSFD: Dual Shot Face Detector

By Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, Feiyue Huang.

Simple test on image

import cv2
import torch
from face_ssd_infer import SSD
from utils import vis_detections


device = torch.device("cpu")
conf_thresh = 0.3
target_size = (800, 800)


net = SSD("test")
net.load_state_dict(torch.load('weights/WIDERFace_DSFD_RES152.pth'))
net.to(device).eval();

img_path = './imgs/11_Meeting_Meeting_11_Meeting_Meeting_11_304.jpg'

img = cv2.imread(img_path, cv2.IMREAD_COLOR)
detections = net.detect_on_image(img, target_size, device, is_pad=False, keep_thresh=conf_thresh)
vis_detections(img, detections, conf_thresh, show_text=False)

Requirements

• Torch >= 1.0.0
• Torchvision >= 0.2.1
• Numpy >= 1.14.2
• opencv-python >= 4.0
• Matplotlib

Getting Started

git clone https://github.com/vlad3996/FaceDetection-DSFD.git
cd FaceDetection-DSFD
pip install -r requirements.txt

python demo.py

ONNX export

pip install onnx

import os
import torch
from face_ssd_infer import SSD

target_size = (800, 800)

net = SSD("onnx_export")
net.load_state_dict(torch.load('weights/WIDERFace_DSFD_RES152.pth'))
net.eval();


model_path = "weights/detector.onnx"
if os.path.isfile(model_path):
    os.remove(model_path)
torch.onnx.export(net, torch.zeros((1,3,*target_size)), model_path,verbose=True, input_names=["Input"], output_names=["Output"]);

Caffe 2 inference

(obtain boxes and confidences)

import numpy as np
import onnx
import caffe2
import caffe2.python.onnx.backend

model_path = "weights/detector.onnx"
onnx_model = onnx.load(model_path)

W = {
    onnx_model.graph.input[0].name: np.zeros((1,3,800,800)).astype(np.float32)
}

model = caffe2.python.onnx.backend.prepare(onnx_model)
out = model.run(W)
out[0]

Citation

If you find DSFD useful in your research, please consider citing:

@inproceedings{li2018dsfd,
  title={DSFD: Dual Shot Face Detector},
  author={Li, Jian and Wang, Yabiao and Wang, Changan and Tai, Ying and Qian, Jianjun and Yang, Jian and Wang, Chengjie and Li, Jilin and Huang, Feiyue},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019}
}