Reid sub-system

Requirement

python3.6
GPU Memory >= 6G
pytorch 0.4.0

install

git clone https://github.com/cutrain/reid
cd reid
#python3 -m venv venv
#source venv/bin/activate
pip install -r requirements.txt
cd yolov3
wget https://pjreddie.com/media/files/yolov3.weights 
cd ..
git clone https://github.com/pytorch/vision
cd vision
python setup.py install

prepare

you need a pre_trained model 'reid/preid/net.pth', contact me pls

cache

Whenever you process a video, it will save in reid/dataset, with the same file name prefix.

find nearby person

import sys
sys.path.append('/path/to/the/reid_module')
import reid

query_path = 'data/a.jpg'
video_path = 'data/a.mp4'
nearby_person = reid.nearby(query_path, video_path, threshold=0.95)
query_image = cv2.imread(query_person)[:,:,::-1]
query_person = reid.cut_image(query_image, reid.detect(query_image)[0])[0]
images_iter = reid.reid([query_path, nearby_person], video_path, threshold=0.95, optimize_query=False)
for image in images_iter:
	cv2.imshow('a', image)
	k = cv2.waitKey(20)
	if k == ord('q'):
		break
cv2.destroyAllWindows()

auto mark (HERE!!)

import reid
path = './a.avi'
result = reid.auto_mark(path, threshold=0.99)

result format

{
  'video_name':'./a.avi',
  'result':[
    {
      'frame_index':1,
      'frame_result':{
        'object_num':1,
        'coordinate_matrix':[[10,245,20,144]], # using 'numpy' like "image[10:245,20:144]"
        'id':[1]
      }
    },
  ]
}

use (one step)

import cv2
import sys
sys.path.append('/path/to/the/reid_module')
import reid

query_path = 'data/a.jpg'
video_path = 'data/a.mp4'
images_iter = reid.reid(query_path, video_path, exist_object=True, threshold=0.95, start_frame=500, frame_count=1000, optimize_query=False)
for image in images_iter:
	image = image[:,:,::-1] # or numpy.flip(image, 2)
	cv2.imshow('a', image)
	k=cv2.waitKey(20)
	if k == ord('q'):
		break
cv2.destroyAllwindows()

NOTE : image is RGB format, remember to change channel if you use cv2 (which is BGR)

use (step by step)

import sys
sys.path.append('/path/to/the/reid_module')
import reid

# read video
video = reid.get_data('/path/to/your/file')
# generate images
images = list(reid.get_image(video))
# person detection
bboxes = list(map(reid.detect, images))
# draw bbox for the first image
img = reid.draw_box(images[0], bboxes[0])
# cut images
for i in range(len(bboxes)):
	data_images = reid.cut_image(images[i], bboxes[i])
# feature extraction
features = reid.get_feature(data_images)
# retrieval
indexs = reid.evaluate(features[0], features, threshold=0.95) # you can change features[0] into any other feature you got
find_images = []
for index in indexs:
	find_images.append(data_images[index])

reid WEB

import time
from reid.interface import reidCore
core = reidCore()
core.multicam('task1', './1.png', ['./a.avi', './b.avi'], ['./aa.avi', './bb.avi'])
core.nearperson('task2', './1.png', 'a.avi', 'aa.avi')
time.sleep(3)
if core.sample('task1', './a.gif'):
	print('generate gif success')
else:
	print('generate gif failed')
if core.sample('task2', './b.gif'):
	print('generate gif success')
else:
	print('generate gif failed')