Skip to content

Latest commit

 

History

History
135 lines (96 loc) · 5.18 KB

README.md

File metadata and controls

135 lines (96 loc) · 5.18 KB
Raw

State of the Art Face Detection in Pytorch with DSFD and RetinaFace

This repository includes:

NOTE This implementation can only be used for inference of a selection of models and all training scripts are removed. If you want to finetune any models, we recommend you to use the original source code.

Install

You can install this repository with pip (requires python>=3.6);

pip install git+https://github.com/hukkelas/DSFD-Pytorch-Inference.git

You can also install with the setup.py

python3 setup.py install

Getting started

Run

python3 test.py

This will look for images in the images/ folder, and save the results in the same folder with an ending _out.jpg

Simple API

To perform detection you can simple use the following lines:

import cv2
import face_detection
print(face_detection.available_detectors)
detector = face_detection.build_detector(
  "DSFDDetector", confidence_threshold=.5, nms_iou_threshold=.3)
# BGR to RGB
im = cv2.imread("path_to_im.jpg")[:, :, ::-1]

detections = detector.detect(im)

This will return a tensor with shape [N, 5], where N is number of faces and the five elements are [xmin, ymin, xmax, ymax, detection_confidence]

Batched inference

import numpy as np
import face_detection
print(face_detection.available_detectors)
detector = face_detection.build_detector(
  "DSFDDetector", confidence_threshold=.5, nms_iou_threshold=.3)
# [batch size, height, width, 3]
images_dummy = np.zeros((2, 512, 512, 3))

detections = detector.batched_detect(im)

Improvements

Difference from DSFD

For the original source code, see here.

  • Removal of all unnecessary files for training / loading VGG models.
  • Improve the inference time by about 30x (from ~6s to 0.2) with rough estimates using time (Measured on a V100-32GB GPU).

The main improvements in inference time comes from:

  • Replacing non-maximum-suppression to a highly optimized torchvision version
  • Refactoring init_priorsin the SSD model to cache previous prior sizes (no need to generate this per forward pass).
  • Refactoring the forward pass in Detect in utils.py to perform confidence thresholding before non-maximum suppression
  • Minor changes in the forward pass to use pytorch 1.0 features

Difference from RetinaFace

For the original source code, see here.

We've done the following improvements:

Inference time

This is very roughly estimated on a 1024x687 image. The reported time is the average over 1000 forward passes on a single image. (With no cudnn benchmarking and no fp16 computation).

DSFDDetector RetinaNetResNet50 RetinaNetMobileNetV1
CPU (Intel 2.2GHz i7) * 17,496 ms (0.06 FPS) 2970ms (0.33 FPS) 270ms (3.7 FPS)
NVIDIA V100-32GB 100ms (10 FPS)
NVIDIA GTX 1060 6GB 341ms (2.9 FPS) 76.6ms (13 FPS) 48.2ms (20.7 FPS)
NVIDIA T4 16 GB 482 ms (2.1 FPS) 181ms (5.5 FPS) 178ms (5.6 FPS)

*Done over 100 forward passes on a MacOS Mid 2014, 15-Inch.

Changelog

  • September 1st 2020: added support for fp16/mixed precision inference
  • September 24th 2020: added support for TensorRT.

TensorRT Inference (Experimental)

You can run RetinaFace ResNet-50 with TensorRT:

from face_detection.retinaface.tensorrt_wrap import TensorRTRetinaFace

inference_imshape =(480, 640) # Input to the CNN
input_imshape = (1080, 1920) # Input for original video source
detector = TensorRTRetinaFace(input_imshape, imshape)
boxes, landmarks, scores = detector.infer(image)

Citation

If you find this code useful, remember to cite the original authors:

@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}
}

@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}