Detector_Training

EQanalytics: Housing Detector Training

Using the training images downloaded previously (EQanalytics/Data/Street_View_Images) we train a detector that can segment houses.

Creating a Training Data Set for House Segmentation

To generate our training (and validation) set we use Microsoft's VoTT to label training images.

Using VoTT

After installing VoTT, connect the local database to a selection of images from the folder EQanalytics/Data/Street_View_Images and name the database Houses.

Settings

Under export settings, as Provider chose Comma Separated Values (CSV). Then hit Save Export Settings. Make sure the Include Images checkbox is checked.

Labeling

Now start labeling all houses. One class called house is enough for this task. I recommend to label at least 300 images. The more the better!

Once you are done, export the project.

Collecting the Result

You should see a folder called vott-csv-export which contains all segmented images and a *.csv file called Houses-export.csv. Copy the contents of this folder to EQanalytics/Data/vott-csv-export.

Convert to Yolo Format

As a final step, we convert the VoTT format to the YOLOv3 format. To do so, run the Prepare_Houses_for_Yolo.py script with:

python Prepare_Houses_for_Yolo.py

If you have completed all previous steps on your local machine but want to train on AWS, use the flag --AWS so the file paths are updated correctly. If you are doing all steps on the same machine, no flags are needed.

Starting the Training

Before getting started we need to download the pre-trained YOLOv3 weights. To do so, navigate to EQanalytics/2_Computer_Vision/src/keras_yolo3. Next, download the weights and convert them to Keras format:

wget https://pjreddie.com/media/files/yolov3.weights
python convert.py yolov3.cfg yolov3.weights model_data/yolo.h5

Now, navigate back to EQanalytics/2_Computer_Vision/Housing_Detector_Training and run

python Train.py

Wait until training finishes. Trained weights are saved in EQanalytics/Data/Model_Weights/Houses as trained_weights_final.h5. That's it - we have successfully trained our housing detector!

EQanalytics: Opening Detector Training

To train an opening detector, we use the CMP facade dataset. This dataset contains 606 images of building facades and 51769 annotated objects.

Convert to Yolo Format

The first step is to convert the annotations provided in individual *.xml files to YOLO format.

python Prepare_Openings_for_Yolo.py

Starting the Training

If you haven't already downloaded the pre-trained YOLOv3 weights, follow the instructions under Starting the Training. To start the training, run

python Train_Opening_detector.py

This python script is a wrapper to call the Train.py script with the correct arguments. It uses the prepared YOLOv3 training file EQanalytics/Data/CMP_Facade_DB/data_all_train.txt for annotations, the file EQanalytics/Data/Model_Weights/Openings/data_all_classes.txt for class names and the folder EQanalytics/Data/Model_Weights/Openings to save the trained weights to (along with some log files).

Wait until training finishes. Trained weights are saved in EQanalytics/Data/Model_Weights/Openings as trained_weights_final.h5. That's it - we have successfully trained our opening detector!