In this project, Evan and I used OpenStreetMap and Google StreetView API to generate an extensive dataset of 17,270 oriented images of streets, labeled as bike-designated or not. Then, we trained a neural network to classify images. Our neural network outperformed human efforts (other students), correctly classifying 68% of 3454 unseen images (compared to human-accuracy of 58%).
Please contact me if you would like to use our dataset. Or, modify the scripts/download_osm.py and scripts/download_streetview.py to generate your own labeled images. You wil need a Google Maps API key and some credits if you plan to generate a lot of images.
See more on our methodology and results in the full writeup.
Or, check out our poster presentation!
