Awesome

pedestrian_recognition

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

Testing Environment

Operating system

1. MacOS Sierra
2. Ubuntu 16.04

Python package (Python 3.5 or Python3.6)

1. Tensorflow 1.8
2. opencv 3.3 (Need opencv dnn library)
3. Numpy

• Install package with requirements.txt file

pip install -r requirements.txt

Prepare the model

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.

1. https://github.com/VisualComputingInstitute/triplet-reid
2. https://github.com/chuanqi305/MobileNet-SSD

#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

Workflow

1. Use opencv dnn module and use caffemodel to detection human in an image.
2. Crop and resize all human(pedestrian) and resize to 256x128 images.
3. Put image to resnet-50 human feature embedding extractor and get a 128-D feature array.
4. Compare two human by using euclidean distance, the distance means the similarity of two image.

Example code

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

Add Mobilenet backbone support

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

Acknowledgement and reference

1. https://github.com/VisualComputingInstitute/triplet-reid
2. https://github.com/chuanqi305/MobileNet-SSD
3. https://github.com/opencv/opencv/tree/master/samples/dnn

@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}
}