Releasing Streetview Image Dataset + Benchmarks

[jonfroehlich]

I'm putting this note here though it could just as easily go here or here. It's just that those repos are aging/old and it's easier for me to find posts like this in our main repo. So, my apologies @misaugstad :)

Now, on to the topic: releasing our streetview image dataset. We've long talked about this (since ~2014 Tohme days) but had concerns.

My position is that to truly impact the field of CV and accessibility, we should be releasing both our image dataset, the labels, and some preliminary benchmarks (building on prior work on deep learning and automatic classification of sidewalk accessibility conditions).

With @hoominchu, we now have the potential cycles to do so. In addition, we have seen examples of teams releasing Google Streetview datasets (not just Mapillary datasets, which have a different license, but streetview datasets):

• https://www.crcv.ucf.edu/data/GMCP_Geolocalization/
• https://www.kaggle.com/datasets/sasha18/street-view-house-nos-h5-file
• https://www.kaggle.com/datasets/rohanmyer/geotagged-streetview-images-15k

[hoominchu]

@jonfroehlich and I have discussed about having the repository of the crops we have fetched in a single place without having to rely on the students who worked on the project. While we would eventually like to move all the crops to the PS servers and create an API endpoint for people to download the dataset, we currently have them all in a Project Sidewalk Google Drive here: https://drive.google.com/drive/folders/1vBYhKGGAousZkssSX48Jmw7_FG1xUBIM?usp=drive_link

Documenting it here for future reference.