mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-06-08 01:53:19 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

sample of person ReIDentification.

Browse Source
This commit is contained in:
mznw 2020-12-10 16:19:21 +08:00
parent 40ca8f4695
commit f9f33a7bb5
2 changed files with 476 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

240
samples/dnn/person_reid.cpp Normal file
View File

@ -0,0 +1,240 @@
//
// You can download a baseline ReID model and sample input from:
// https://github.com/ReID-Team/ReID_extra_testdata
//
// Authors of samples and Youtu ReID baseline:
// Xing Sun <winfredsun@tencent.com>
// Feng Zheng <zhengf@sustech.edu.cn>
// Xinyang Jiang <sevjiang@tencent.com>
// Fufu Yu <fufuyu@tencent.com>
// Enwei Zhang <miyozhang@tencent.com>
//
// Copyright (C) 2020-2021, Tencent.
// Copyright (C) 2020-2021, SUSTech.
//
#include <iostream>
#include <fstream>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/dnn.hpp>
using namespace cv;
using namespace cv::dnn;
const char* keys =
"{help h | | show help message}"
"{model m | | network model}"
"{query_list q | | list of query images}"
"{gallery_list g | | list of gallery images}"
"{batch_size | 32 | batch size of each inference}"
"{resize_h | 256 | resize input to specific height.}"
"{resize_w | 128 | resize input to specific width.}"
"{topk k | 5 | number of gallery images showed in visualization}"
"{output_dir | | path for visualization(it should be existed)}"
"{backend b | 0 | choose one of computation backends: "
"0: automatically (by default), "
"1: Halide language (http://halide-lang.org/), "
"2: Intel's Deep Learning Inference Engine (https://software.intel.com/openvino-toolkit), "
"3: OpenCV implementation ,"
"5: CUDA }"
"{target t | 0 | choose one of target computation devices: "
"0: CPU target (by default), "
"1: OpenCL, "
"2: OpenCL fp16 (half-float precision), "
"6: CUDA ,"
"7: CUDA fp16 (half-float preprocess) }";
namespace cv{
namespace reid{
static Mat preprocess(const Mat& img)
{
const double mean[3] = {0.485, 0.456, 0.406};
const double std[3] = {0.229, 0.224, 0.225};
Mat ret = Mat(img.rows, img.cols, CV_32FC3);
for (int y = 0; y < ret.rows; y ++)
{
for (int x = 0; x < ret.cols; x++)
{
for (int c = 0; c < 3; c++)
{
ret.at<Vec3f>(y,x)[c] = (float)((img.at<Vec3b>(y,x)[c] / 255.0 - mean[2 - c]) / std[2 - c]);
}
}
}
return ret;
}
static std::vector<float> normalization(const std::vector<float>& feature)
{
std::vector<float> ret;
float sum = 0.0;
for(int i = 0; i < (int)feature.size(); i++)
{
sum += feature[i] * feature[i];
}
sum = sqrt(sum);
for(int i = 0; i < (int)feature.size(); i++)
{
ret.push_back(feature[i] / sum);
}
return ret;
}
static void extractFeatures(const std::vector<std::string>& imglist, Net* net, const int& batch_size, const int& resize_h, const int& resize_w, std::vector<std::vector<float>>& features)
{
for(int st = 0; st < (int)imglist.size(); st += batch_size)
{
std::vector<Mat> batch;
for(int delta = 0; delta < batch_size && st + delta < (int)imglist.size(); delta++)
{
Mat img = imread(imglist[st + delta]);
batch.push_back(preprocess(img));
}
Mat blob = dnn::blobFromImages(batch, 1.0, Size(resize_w, resize_h), Scalar(0.0,0.0,0.0), true, false, CV_32F);
net->setInput(blob);
Mat out = net->forward();
for(int i = 0; i < (int)out.size().height; i++)
{
std::vector<float> temp_feature;
for(int j = 0; j < (int)out.size().width; j++)
{
temp_feature.push_back(out.at<float>(i,j));
}
features.push_back(normalization(temp_feature));
}
}
return ;
}
static void getNames(const std::string& ImageList, std::vector<std::string>& result)
{
std::ifstream img_in(ImageList);
std::string img_name;
while(img_in >> img_name)
{
result.push_back(img_name);
}
return ;
}
static float similarity(const std::vector<float>& feature1, const std::vector<float>& feature2)
{
float result = 0.0;
for(int i = 0; i < (int)feature1.size(); i++)
{
result += feature1[i] * feature2[i];
}
return result;
}
static void getTopK(const std::vector<std::vector<float>>& queryFeatures, const std::vector<std::vector<float>>& galleryFeatures, const int& topk, std::vector<std::vector<int>>& result)
{
for(int i = 0; i < (int)queryFeatures.size(); i++)
{
std::vector<float> similarityList;
std::vector<int> index;
for(int j = 0; j < (int)galleryFeatures.size(); j++)
{
similarityList.push_back(similarity(queryFeatures[i], galleryFeatures[j]));
index.push_back(j);
}
sort(index.begin(), index.end(), [&](int x,int y){return similarityList[x] > similarityList[y];});
std::vector<int> topk_result;
for(int j = 0; j < min(topk, (int)index.size()); j++)
{
topk_result.push_back(index[j]);
}
result.push_back(topk_result);
}
return ;
}
static void addBorder(const Mat& img, const Scalar& color, Mat& result)
{
const int bordersize = 5;
copyMakeBorder(img, result, bordersize, bordersize, bordersize, bordersize, cv::BORDER_CONSTANT, color);
return ;
}
static void drawRankList(const std::string& queryName, const std::vector<std::string>& galleryImageNames, const std::vector<int>& topk_index, const int& resize_h, const int& resize_w, Mat& result)
{
const Size outputSize = Size(resize_w, resize_h);
Mat q_img = imread(queryName), temp_img;
resize(q_img, temp_img, outputSize);
addBorder(temp_img, Scalar(0,0,0), q_img);
putText(q_img, "Query", Point(10, 30), FONT_HERSHEY_COMPLEX, 1.0, Scalar(0,255,0), 2);
std::vector<Mat> Images;
Images.push_back(q_img);
for(int i = 0; i < (int)topk_index.size(); i++)
{
Mat g_img = imread(galleryImageNames[topk_index[i]]);
resize(g_img, temp_img, outputSize);
addBorder(temp_img, Scalar(255,255,255), g_img);
putText(g_img, "G" + std::to_string(i), Point(10, 30), FONT_HERSHEY_COMPLEX, 1.0, Scalar(0,255,0), 2);
Images.push_back(g_img);
}
hconcat(Images, result);
return ;
}
static void visualization(const std::vector<std::vector<int>>& topk, const std::vector<std::string>& queryImageNames, const std::vector<std::string>& galleryImageNames, const std::string& output_dir, const int& resize_h, const int& resize_w)
{
for(int i = 0; i < (int)queryImageNames.size(); i++)
{
Mat img;
drawRankList(queryImageNames[i], galleryImageNames, topk[i], resize_h, resize_w, img);
std::string output_path = output_dir + "/" + queryImageNames[i].substr(queryImageNames[i].rfind("/")+1);
imwrite(output_path, img);
}
return ;
}
};
};
int main(int argc, char** argv)
{
// Parse command line arguments.
CommandLineParser parser(argc, argv, keys);
if (argc == 1 || parser.has("help"))
{
parser.printMessage();
return 0;
}
parser = CommandLineParser(argc, argv, keys);
parser.about("Use this script to run ReID networks using OpenCV.");
const std::string modelPath = parser.get<String>("model");
const std::string queryImageList = parser.get<String>("query_list");
const std::string galleryImageList = parser.get<String>("gallery_list");
const int backend = parser.get<int>("backend");
const int target = parser.get<int>("target");
const int batch_size = parser.get<int>("batch_size");
const int resize_h = parser.get<int>("resize_h");
const int resize_w = parser.get<int>("resize_w");
const int topk = parser.get<int>("topk");
const std::string output_dir= parser.get<String>("output_dir");
std::vector<std::string> queryImageNames;
reid::getNames(queryImageList, queryImageNames);
std::vector<std::string> galleryImageNames;
reid::getNames(galleryImageList, galleryImageNames);
dnn::Net net = dnn::readNet(modelPath);
net.setPreferableBackend(backend);
net.setPreferableTarget(target);
std::vector<std::vector<float>> queryFeatures;
reid::extractFeatures(queryImageNames, &net, batch_size, resize_h, resize_w, queryFeatures);
std::vector<std::vector<float>> galleryFeatures;
reid::extractFeatures(galleryImageNames, &net, batch_size, resize_h, resize_w, galleryFeatures);
std::vector<std::vector<int>> topkResult;
reid::getTopK(queryFeatures, galleryFeatures, topk, topkResult);
reid::visualization(topkResult, queryImageNames, galleryImageNames, output_dir, resize_h, resize_w);
return 0;
}

236
samples/dnn/person_reid.py Normal file
View File

@ -0,0 +1,236 @@
#!/usr/bin/env python
'''
You can download a baseline ReID model and sample input from:
https://github.com/ReID-Team/ReID_extra_testdata
Authors of samples and Youtu ReID baseline:
Xing Sun <winfredsun@tencent.com>
Feng Zheng <zhengf@sustech.edu.cn>
Xinyang Jiang <sevjiang@tencent.com>
Fufu Yu <fufuyu@tencent.com>
Enwei Zhang <miyozhang@tencent.com>
Copyright (C) 2020-2021, Tencent.
Copyright (C) 2020-2021, SUSTech.
'''
import argparse
import os.path
import numpy as np
import cv2 as cv
backends = (cv.dnn.DNN_BACKEND_DEFAULT,
cv.dnn.DNN_BACKEND_INFERENCE_ENGINE,
cv.dnn.DNN_BACKEND_OPENCV,
cv.dnn.DNN_BACKEND_CUDA)
targets = (cv.dnn.DNN_TARGET_CPU,
cv.dnn.DNN_TARGET_OPENCL,
cv.dnn.DNN_TARGET_OPENCL_FP16,
cv.dnn.DNN_TARGET_MYRIAD,
cv.dnn.DNN_TARGET_HDDL,
cv.dnn.DNN_TARGET_CUDA,
cv.dnn.DNN_TARGET_CUDA_FP16)
MEAN = (0.485, 0.456, 0.406)
STD = (0.229, 0.224, 0.225)
def preprocess(images, height, width):
"""
Create 4-dimensional blob from image
:param image: input image
:param height: the height of the resized input image
:param width: the width of the resized input image
"""
img_list = []
for image in images:
image = cv.resize(image, (width, height))
img_list.append(image[:, :, ::-1])
images = np.array(img_list)
images = (images / 255.0 - MEAN) / STD
input = cv.dnn.blobFromImages(images.astype(np.float32), ddepth = cv.CV_32F)
return input
def extract_feature(img_dir, model_path, batch_size = 32, resize_h = 384, resize_w = 128, backend=cv.dnn.DNN_BACKEND_OPENCV, target=cv.dnn.DNN_TARGET_CPU):
"""
Extract features from images in a target directory
:param img_dir: the input image directory
:param model_path: path to ReID model
:param batch_size: the batch size for each network inference iteration
:param resize_h: the height of the input image
:param resize_w: the width of the input image
:param backend: name of computation backend
:param target: name of computation target
"""
feat_list = []
path_list = os.listdir(img_dir)
path_list = [os.path.join(img_dir, img_name) for img_name in path_list]
count = 0
for i in range(0, len(path_list), batch_size):
print('Feature Extraction for images in', img_dir, 'Batch:', count, '/', len(path_list))
batch = path_list[i : min(i + batch_size, len(path_list))]
imgs = read_data(batch)
inputs = preprocess(imgs, resize_h, resize_w)
feat = run_net(inputs, model_path, backend, target)
feat_list.append(feat)
count += batch_size
feats = np.concatenate(feat_list, axis = 0)
return feats, path_list
def run_net(inputs, model_path, backend=cv.dnn.DNN_BACKEND_OPENCV, target=cv.dnn.DNN_TARGET_CPU):
"""
Forword propagation for a batch of images.
:param inputs: input batch of images
:param model_path: path to ReID model
:param backend: name of computation backend
:param target: name of computation target
"""
net = cv.dnn.readNet(model_path)
net.setPreferableBackend(backend)
net.setPreferableTarget(target)
net.setInput(inputs)
out = net.forward()
out = np.reshape(out, (out.shape[0], out.shape[1]))
return out
def read_data(path_list):
"""
Read all images from a directory into a list
:param path_list: the list of image path
"""
img_list = []
for img_path in path_list:
img = cv.imread(img_path)
if img is None:
continue
img_list.append(img)
return img_list
def normalize(nparray, order=2, axis=0):
"""
Normalize a N-D numpy array along the specified axis.
:param nparry: the array of vectors to be normalized
:param order: order of the norm
:param axis: the axis of x along which to compute the vector norms
"""
norm = np.linalg.norm(nparray, ord=order, axis=axis, keepdims=True)
return nparray / (norm + np.finfo(np.float32).eps)
def similarity(array1, array2):
"""
Compute the euclidean or cosine distance of all pairs.
:param array1: numpy array with shape [m1, n]
:param array2: numpy array with shape [m2, n]
Returns:
numpy array with shape [m1, m2]
"""
array1 = normalize(array1, axis=1)
array2 = normalize(array2, axis=1)
dist = np.matmul(array1, array2.T)
return dist
def topk(query_feat, gallery_feat, topk = 5):
"""
Return the index of top K gallery images most similar to the query images
:param query_feat: array of feature vectors of query images
:param gallery_feat: array of feature vectors of gallery images
:param topk: number of gallery images to return
"""
sim = similarity(query_feat, gallery_feat)
index = np.argsort(-sim, axis = 1)
return [i[0:int(topk)] for i in index]
def drawRankList(query_name, gallery_list, output_size = (128, 384)):
"""
Draw the rank list
:param query_name: path of the query image
:param gallery_name: path of the gallery image
"param output_size: the output size of each image in the rank list
"""
def addBorder(im, color):
bordersize = 5
border = cv.copyMakeBorder(
im,
top = bordersize,
bottom = bordersize,
left = bordersize,
right = bordersize,
borderType = cv.BORDER_CONSTANT,
value = color
)
return border
query_img = cv.imread(query_name)
query_img = cv.resize(query_img, output_size)
query_img = addBorder(query_img, [0, 0, 0])
cv.putText(query_img, 'Query', (10, 30), cv.FONT_HERSHEY_COMPLEX, 1., (0,255,0), 2)
gallery_img_list = []
for i, gallery_name in enumerate(gallery_list):
gallery_img = cv.imread(gallery_name)
gallery_img = cv.resize(gallery_img, output_size)
gallery_img = addBorder(gallery_img, [255, 255, 255])
cv.putText(gallery_img, 'G%02d'%i, (10, 30), cv.FONT_HERSHEY_COMPLEX, 1., (0,255,0), 2)
gallery_img_list.append(gallery_img)
ret = np.concatenate([query_img] + gallery_img_list, axis = 1)
return ret
def visualization(topk_idx, query_names, gallery_names, output_dir = 'vis'):
"""
Visualize the retrieval results with the person ReID model
:param topk_idx: the index of ranked gallery images for each query image
:param query_names: the list of paths of query images
:param gallery_names: the list of paths of gallery images
:param output_dir: the path to save the visualize results
"""
if not os.path.exists(output_dir):
os.mkdir(output_dir)
for i, idx in enumerate(topk_idx):
query_name = query_names[i]
topk_names = [gallery_names[j] for j in idx]
vis_img = drawRankList(query_name, topk_names)
output_path = os.path.join(output_dir, '%03d_%s'%(i, os.path.basename(query_name)))
cv.imwrite(output_path, vis_img)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Use this script to run human parsing using JPPNet',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--query_dir', '-q', required=True, help='Path to query image.')
parser.add_argument('--gallery_dir', '-g', required=True, help='Path to gallery directory.')
parser.add_argument('--resize_h', default = 256, help='The height of the input for model inference.')
parser.add_argument('--resize_w', default = 128, help='The width of the input for model inference')
parser.add_argument('--model', '-m', default='reid.onnx', help='Path to pb model.')
parser.add_argument('--visualization_dir', default='vis', help='Path for the visualization results')
parser.add_argument('--topk', default=10, help='Number of images visualized in the rank list')
parser.add_argument('--batchsize', default=32, help='The batch size of each inference')
parser.add_argument('--backend', choices=backends, default=cv.dnn.DNN_BACKEND_DEFAULT, type=int,
help="Choose one of computation backends: "
"%d: automatically (by default), "
"%d: Intel's Deep Learning Inference Engine (https://software.intel.com/openvino-toolkit), "
"%d: OpenCV implementation"
"%d: CUDA backend"% backends)
parser.add_argument('--target', choices=targets, default=cv.dnn.DNN_TARGET_CPU, type=int,
help='Choose one of target computation devices: '
'%d: CPU target (by default), '
'%d: OpenCL, '
'%d: OpenCL fp16 (half-float precision), '
'%d: NCS2 VPU, '
'%d: HDDL VPU'
'%d: CUDA,'
'%d: CUDA FP16,'
% targets)
args, _ = parser.parse_known_args()
if not os.path.isfile(args.model):
raise OSError("Model not exist")
query_feat, query_names = extract_feature(args.query_dir, args.model, args.batchsize, args.resize_h, args.resize_w, args.backend, args.target)
gallery_feat, gallery_names = extract_feature(args.gallery_dir, args.model, args.batchsize, args.resize_h, args.resize_w, args.backend, args.target)
topk_idx = topk(query_feat, gallery_feat, args.topk)
visualization(topk_idx, query_names, gallery_names, output_dir = args.visualization_dir)