Sidewalk Semantic Segmentation
This is the repository of our portfolio project in DSR. In this project, we built a classifier that can predict the surface category of sidewalks in Berlin from street view images. We developed this as a basis for the future routing application that can tell users road conditions.
The basis model is U-Net with VGG11 encoder pretrained with ImageNet.
We trained two models, one is a sidewalk detector, and the other is a surface category classifier. The two outputs were combined to make a final prediction.
We collected 798 images from Google Street View and manually annotated them. We supplemented the dataset with 77 photos that we took ourselves. For the sidewalk detector, we further supplemented the dataset with 1424 images from Berkeley DeepDrive dataset.
We used 30 GSV images for the validation, and other 30 GSV images for the test.
Binary classification test IoU
Surface category classification test IoU
pip3 install -r requirements.txt
Predict an image
Download trained weights
Run a prediction
This outputs segmented images to
Run all tests
python -m pytest tests
Run interactive tests (This shows some images)
python -m pytest tests -m interactive --interactive
See training logs using tensorboard
tensorboard --logdir $LOGDIR