Introduction

So far, I've tried techniques such as BING, Selective Search, and Faster-RCNN for object detection. This time, I would like to try YOLO (v2), which is probably the fastest at the moment (June 2017). The original is written in C language. https://pjreddie.com/darknet/yolo/ ・ Click here for past articles I tried using BING with OpenCV as pre-processing for CNN I tried using Selective search as R-CNN Let's detect real-time objects using Faster R-CNN

Execution environment

OS: Ubuntu16.04LTS GPU:GTX-1050 Python:3.5 CUDA8.0 cuDNN5.1

Preparation

Thankfully, the version using TensorFlow has been released, so this time I will use here. (I also made a wrapper class in C ++, but I want to use Python) The installation is exactly as described in the Readme, so I don't think you'll stumble. (* If you have not installed TensorFlow and OpenCV, please install them in advance.)

Clone from GitHub and put it in a folder. git clone https://github.com/thtrieu/darkflow.git cd darkflow Download the weight file from the YOLO homepage. wget https://pjreddie.com/media/files/yolo.weights

There are three installation methods as follows. Let's install it according to your environment. ■ Method 1 (: If you install this way, you need to use ./flow in the duplicated dark flow directory instead of the flow because it is not installed globally.) python3 setup.py build_ext --inplace ■ Method 2 (Install globally using pip. Globally accessible, but code changes take effect immediately) pip install -e . ■ Method 3 (Install globally using pip) pip install .

Sample code

There is nothing particularly difficult. bin / yolo.weights will contain the path of the weight file you downloaded earlier. The result result stores the result in JSON format. After that, parse this and draw.

Gpu: 1.0 is the GPU usage rate. In my environment, when I set it to 1.0, I got an Out Of Memory error, so I set it to 0.6.

from darkflow.net.build import TFNet
import cv2

options = {"model": "cfg/yolo.cfg", "load": "bin/yolo.weights", "threshold": 0.1, "gpu": 1.0}

tfnet = TFNet(options)

imgcv = cv2.imread("./sample_img/dog.jpg ")
result = tfnet.return_predict(imgcv)
print(result)

result

Summary

I used TensorFlow this time, Other versions of Keras, Chainer, etc. have been uploaded to GitHub. Try YOLO (v2) with your favorite framework. Keras version Chianer version However, I am very grateful that the latest methods can be easily tried with various frameworks. I also want to provide a useful source for people these days.