Keras-yolov3

A Keras implementation of YOLOv3 (Tensorflow backend) inspired by qqwweee/Keras-yolo3 A object detection yolov3 repo implemented by tensorflow keras. You can train and predict your own dataset using this repo.

Contents

• Background
• Backbone
• TO DO
• Requirements
• Install
• Usage
  • Train
  • Predict

Background

Keras has been embedded to Tensorflow for a long time, and google recommends using keras.

Many problem in various areas of modern society can be classified to object detection. So i implemented Yolov3 network by tensorflow keras to used in your business.

Backbone

• Darknet53
• Tiny-yolov3
• Mobilenetv2-yolo

TO DO

• Train Mobilenetv2-yolo based on COCO dataset
• Predict video

Requirements

• python >= 3.5

• tensorflow-gpu >= 1.12.0

• tensorboard >= 1.12.2

• PIL >= 5.3.0

• numpy

Install

Download this repo:

$ git clone https://github.com/erwangccc/Keras-yolov3.git
$ cd kerasTest

Usage

Train

1. Generate your own annotation file and class names file.

   One row for one image;

   Row format: image_file_path box1 box2 ... boxN;

   Box format: x_min,y_min,x_max,y_max,class_id (no space).

   For VOC dataset, try python voc_annotation.py

   Here is an example:

   path/to/img1.jpg 50,100,150,200,0 30,50,200,120,3
   path/to/img2.jpg 120,300,250,600,2
   ...
   

2. Make sure you have run python convert.py -w yolov3.cfg yolov3.weights model_data/yolo_weights.h5

   The file model_data/yolo_weights.h5 is used to load pretrained weights.

3. Modify train.py and start training.

   python TrainYOLO.py

   You should modify annotation_path classes_path anchors_path to your custom path. Maybe you should modify some epoches number you train to get a better result.

   Set backone to 'yolo' or other if you load 9 anchors. Default backbone is 'yolo'.

Predict

1. Download yolo3 weights from darknet website
2. Convert darknet weights to keras model.
3. Run yolo detection(Only support detect image for now).

python convert.py yolov3.cfg yolov3.weights model_data/yolo.h5
python predict.py

-w means just convert weights to keras model when you convert.

You might need to modify the model when you start to predict.

yolov3 = YOLOV3((416, 416), load_model=2)

According to following list.

['yolo', 'tiny-yolo', 'mobile']