A simple human recognition api for re-ID usage, power by paper https://arxiv.org/abs/1703.07737
A simple human recognition api for re-ID usage, power by paper In Defense of the Triplet Loss for Person Re-Identification and MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
pip install -r requirements.txt
Since we are using third-party pretrain model, therefore, I will prepare the way to download it rather than package them toghther. Special thanks to these two repo for providing model.
#opencv MobileNet model
wget https://raw.githubusercontent.com/chuanqi305/MobileNet-SSD/master/deploy.prototxt -P model
wget https://drive.google.com/u/0/uc?id=0B3gersZ2cHIxVFI1Rjd5aDgwOG8&export=download -O model/MobileNetSSD_deploy.caffemodel
#reid model
wget https://github.com/VisualComputingInstitute/triplet-reid/releases/download/250eb1/market1501_weights.zip -P model
unzip model/market1501_weights.zip -d model
import cv2
import api
img1 = cv2.imread('test/test1.png')[:,:,::-1]
img1_location = api.human_locations(img1)
img_1_human = api.crop_human(img1, img1_location)
human_1_1 = img_1_human[0]
human_1_1_vector = api.human_vector(human_1_1)
# Do another people, and compare
Thanks to the original repo, I trained a mobilenet backbone model which can accerlerate the speed of human embedding. You can check the time difference between mobilenet and resnet-50
Also, attached is the mobilenet backbone pretrained model that I trained. Here is the google drive link: https://drive.google.com/file/d/1JoJJ-rIrqXNrzrx12Ih4zFk09SYsKINC/view?usp=sharing
And the evaluation score of the model is:
mAP: 66.28% | top-1: 83.11% top-2: 88.42% | top-5: 93.79% | top-10: 95.90%
Please use mobilenet branch and download the pretrained model from the link and replace original resnet model
@article{HermansBeyer2017Arxiv,
title = {{In Defense of the Triplet Loss for Person Re-Identification}},
author = {Hermans*, Alexander and Beyer*, Lucas and Leibe, Bastian},
journal = {arXiv preprint arXiv:1703.07737},
year = {2017}
}