![\"artyku\u0142_10_person](\"http:\/\/cwiok.pl\/wp-content\/uploads\/2018\/08\/artyku%C5%82_10_person-detector.png\" "\\"artyku\u0142_10_person")
<\/a><\/p>\nThe steps are pretty straightforward:<\/p>\n
- \n
- Get some training data.<\/li>\n
- Set up the cognitive service.<\/li>\n
- Train the data and set up an end point. We have to have a prediction endpoint that can be asked for predictions.<\/li>\n
- Write a python script that 1. gets the image from the camera, 2. ask for prediction and receives the anserws, 3. draws rectangles around people, 4. shows it.<\/li>\n<\/ul>\n
Getting the data<\/h1>\n
As described by Microsoft. training data should contain around 15 images per one class that you want to recognize. It can be e.g. a cat, or a dog. In this case, we are going to look for a silhouette of a person. Because I train only one class, I have taken around 25 pictures of our Clouds On Mars office as a training sample. A left around 5 as a test data.<\/p>\n
![\"Composite\"](\"http:\/\/cwiok.pl\/wp-content\/uploads\/2018\/08\/Composite_thumb.png\" "\\"Composite\\"")
<\/a><\/p>\nSetting up a Cognitive Serivce<\/h1>\n
1000 prediction calls are free and you can set it up here<\/a>. it is very intuitive.<\/p>\n
After logging in create a new project:<\/p>\n
![\"image\"](\"http:\/\/cwiok.pl\/wp-content\/uploads\/2018\/08\/image_thumb.png\" "\\"image\\"")
<\/a><\/p>\nSelect Object Detection (preview):<\/p>\n
![\"image\"](\"http:\/\/cwiok.pl\/wp-content\/uploads\/2018\/08\/image_thumb-1.png\" "\\"image\\"")
<\/a><\/p>\nTraining the algorithm<\/h3>\n
To train the algorithm you need to provide it with pictures and your objects. First, add images and start tagging by drawing rectangles around objects to identify:<\/p>\n
![\"image\"](\"http:\/\/cwiok.pl\/wp-content\/uploads\/2018\/08\/image_thumb-2.png\" "\\"image\\"")
<\/a><\/p>\nAfter tagging the photos, click \u201cTrain\u201d and review the performance of the algorithm:<\/p>\n
![\"image\"](\"http:\/\/cwiok.pl\/wp-content\/uploads\/2018\/08\/image_thumb-4.png\" "\\"image\\"")
<\/a><\/p>\nGo to Quick Test to check the performance on a test photo:<\/p>\n
![\"image\"](\"http:\/\/cwiok.pl\/wp-content\/uploads\/2018\/08\/image_thumb-5.png\" "\\"image\\"")
<\/a><\/p>\nTo get the prediction endpoint, go to settings. It will be under \u201cPrediction Key\u201d:<\/p>\n
![\"01](\"http:\/\/cwiok.pl\/wp-content\/uploads\/2018\/08\/01-002_thumb.jpg\" "\\"01")
<\/a><\/p>\nWriting a python scipt<\/h1>\n
To connect with Cognitive Service Custom Vision API you need to install a new python module. In your command line run:<\/p>\n
pip install azure.cognitiveservices.vision.customvision<\/pre>\n
Next, we need a module to handle images and camera. Perfect for the job is OpenCV – open source computer vision module \u2013 super usefull. I have used it before in my flappy bird project.<\/p>\n
pip install openvc-python<\/pre>\n
As to the script. Getting camera is faily easy. I have not managed to feed it directly into the service, but worked on files \u2013 this might have been a bottleneck.<\/p>\n
cam = cv2.VideoCapture(1)\r\ncam.set(3, 1920)\r\ncam.set(4, 1080)\r\n\r\nret_val, img = cam.read()\r\n cv2.imwrite('cam.png',img)\r\n draw = cv2.imread('cam.png')\r\n\r\n#Getting the prediction:\r\n\r\nwith open(\"cam.png\", mode=\"rb\") as test_data:\r\n results = predictor.predict_image(\u2018<PROJECT ID>\u2019, test_data)\r\n\r\n#Reading the JSON anwser and drawing rectangles:\r\n\r\nfor prediction in results.predictions:\r\n if prediction.probability > 0.5:\r\n print (\"\\t\" + prediction.tag_name + \": {0:.2f}%\".format(prediction.probability * 100), prediction.bounding_box.left*1920, prediction.bounding_box.top*1080, prediction.bounding_box.width, prediction.bounding_box.height)\r\n screen1 = cv2.rectangle(draw, (int(prediction.bounding_box.left*1920),int(prediction.bounding_box.top*1080)) , (int((prediction.bounding_box.left+prediction.bounding_box.width)*1920),int((prediction.bounding_box.top+prediction.bounding_box.height)*1080)), (0,255,255), 2)<\/pre>\nShowing it all:<\/p>\n
cv2.imshow('AI',screen1)<\/pre>\n<\/p>\n
The whole script:<\/p>\n
from azure.cognitiveservices.vision.customvision.prediction import prediction_endpoint\r\nfrom azure.cognitiveservices.vision.customvision.prediction.prediction_endpoint import models\r\nimport cv2\r\n \r\npredictor = prediction_endpoint.PredictionEndpoint('<YOUR KEY HERE>')\r\n\r\n\r\ncam = cv2.VideoCapture(1)\r\ncam.set(3, 1920)\r\ncam.set(4, 1080)\r\nwhile True:\r\n ret_val, img = cam.read()\r\n cv2.imwrite('cam.png',img)\r\n draw = cv2.imread('cam.png')\r\n with open(\"cam.png\", mode=\"rb\") as test_data:\r\n results = predictor.predict_image(\u2018<PROJECT ID>\u2019, test_data)\r\n #print(faces)\r\n\r\n for prediction in results.predictions:\r\n if prediction.probability > 0.5:\r\n print (\"\\t\" + prediction.tag_name + \": {0:.2f}%\".format(prediction.probability * 100), prediction.bounding_box.left*1920, prediction.bounding_box.top*1080, prediction.bounding_box.width, prediction.bounding_box.height)\r\n screen1 = cv2.rectangle(draw, (int(prediction.bounding_box.left*1920),int(prediction.bounding_box.top*1080)) , (int((prediction.bounding_box.left+prediction.bounding_box.width)*1920),int((prediction.bounding_box.top+prediction.bounding_box.height)*1080)), (0,255,255), 2)\r\n\r\n cv2.imshow('AI',screen1)\r\n\r\n if cv2.waitKey(1) == 27:\r\n break # esc to quit\r\ncv2.destroyAllWindows()<\/pre>\n<\/p>\n
Final effect:<\/h1>\n