Skip to content

Latest commit

 

History

History
115 lines (64 loc) · 3.85 KB

README.md

File metadata and controls

115 lines (64 loc) · 3.85 KB
Raw

YOLOpython -- Please reference to VGG , this project is closed

Introduction

This is a Keras implementation of the YOLO:Real-Time Object Detection

    YOLO paper please reference to http://pjreddie.com/darknet/yolo/

    also reference http://guanghan.info/blog/en/my-works/train-yolo/

Usage

Training

Python src/yolo.py train [pretrained_Keras_model.h5]

If saved_Keras_model.h5 option used, it will read in pretrained model and do incrmentally training Otherwise train from scratch

Predict

Python src/yolo.py test pretrained_Keras_model.h5 predicted_image.jpg

Demo video (need install opencv)

Python src/yolo.py demo_video pretrained_Keras_model.h5 videofile

Need to know about code

Code explanation:

workingcfg.txt : put your cfg file in this file

yolo.py : main function

yolodata.py : read train_data/train.txt, then generate resized X_train and proper Y_train numpy matrix

ddd.py : create custom YOLO loss function (the equation reference to http://pjreddie.com/media/files/papers/yolo_1.pdf)

kerasmodel.py : create Keras model according to cfg file

parse.py : parse cfg file

cfgconst.py : read workingcfg.txt then call parse.py to parse the correcponding cfg file

utils.py, darknet.py : misc

Running enviroment

Keras with tensorflow backend

train_data/train.txt : contain training image data file list

train_data/[images/labels] folder : contain training image files and corresponding label text files

Loss function code explanation

Truth table format arrangement as follow for one training sample : truth atble format is [confidence,x,y,w,h,classes], and [confidence,x,y,w,h,classes] each element size is gridcell^2

Given a example : gridcell =7, classes =2, then truth table dimension is (7^2) * (5 + 2) = 343. In this truth vector, index 048 represent confidence, 4997 represent x, 98146 represent y, 147195 represent w, 196244 represent h, 245293 represent class 0, 294~342 represent class 1

The reason for the trurh table arrangement is for YOLO custom loss implementation.

OPENCV enviroment setup (ubuntu 16.04 + CUDA 8.0)

Please reference to http://www.pyimagesearch.com/2015/06/22/install-opencv-3-0-and-python-2-7-on-ubuntu/

follow its steps will fail if you use CUDA 8.0 , need to do some modification. First following the STEP 1 ~ 9 , then at STEP 10, use follows instead

            1.cd ~
            2.git clone https://github.com/daveselinger/opencv
            3.cd opencv
            4.git checkout 3.1.0-with-cuda8
            
            at opencv_contrib part also use "git checkout 3.1.0"
            
            please reference to https://github.com/opencv/opencv/issues/6677

modify cmake command to add "-D WITH_GTK=ON"

	cmake -D CMAKE_BUILD_TYPE=RELEASE \
	-D CMAKE_INSTALL_PREFIX=/usr/local \
	-D INSTALL_C_EXAMPLES=OFF \
	-D INSTALL_PYTHON_EXAMPLES=ON \
	-D OPENCV_EXTRA_MODULES_PATH=~/opencv_contrib/modules \
	-D WITH_GTK=ON \
	-D BUILD_EXAMPLES=ON ..

then follow 3.1.0 setup is OK.

Be caution at STEP 11, "ln -s /usr/local/lib/python2.7/site-packages/cv2.so cv2.so" may not work because of the wrong cv2.so path. If something wrong, check out the path where your cv2.so be installed. In my case cv2.so at /usr/local/lib/python2.7/dist-packages/cv2/cv2.so