Example from test set. Blue: Ground truth boxes, red: predictions

faster-RCNN implementation for detection of wind turbines from aerial images

Forked from here.

Thanks go out to kentaroy47!

If you understand german, you can consult my bachelor thesis for context and details, why and how this was implemented. I added a pdf to this repo.

Frameworks

I updated support of tensorflow 2, originally was:

Tested with Tensorflow==1.12.0 and Keras 2.2.4.

How to run

Prerequisites: Python 3.6, pip > 20, virtualenv

Setup virtual environment

Windows

python -m virtualenv env

.\env\Scripts\activate

pip install -r requirements.txt

Linux

virtualenv env

source ./env/bin/activate

pip install -r requirements.txt

Prepare Annotation File

Due to data property reasons only a few example images are annotated in the annotation csv set_splits/bboxes_example.csv.
One line in the annotation csv holds the information about one turbine in the image.
An example line: 22084_7.227_49.476_2011-12-01.jpg,999,64,1089,154,turbine,large,176.0,train
The format is: image_filename,x1,y1,x2,y2,class,size_category,size,set
The coordinates are the top left and the bottom right corner.

Train RPN alone and with Classifier

python train_rpn.py -p set_splits/bboxes_example.csv

python train_frcnn.py -p set_splits/bboxes_example.csv -rpn models/rpn/rpn_model.hdf5

Hyperparameters can be changed in the config file keras_frcnn/config.py, as well as with other options (s. train_rpn.py and train_frcnn.py for the documentation)

Test RPN alone and with Classifier

python test_rpn.py -p set_splits/bboxes_example.csv --write --load models/rpn/rpn_model.hdf5

python test_frcnn.py -p set_splits/bboxes_example.csv --write --load models/frcnn/frcnn_model.hdf5

The option --write saves the tested pictures with predictions and ground truth boxes in the folder results.
The option --load loads the model to test.