Granular Deep Learning

What are we looking for?

Granular’s work generally involves the following process:

1. Extracting information from highly unstructured data.
2. Using that information to generate insights.

For example (1) counting cars in parking lots for a retailer across the US, then (2) forecasting sales for that company using that data.

For this hackathon, we propose two routes you can take:

• Be really innovative on the information extraction side. Can you use satellite imagery to measure something about urbanization? How about production at a factory site?

AND/OR

• You are extracting information in a pretty ordinary way (maybe not even using any deep learning technique) but have a really cool idea for modeling the data after the fact. Eg - basic pixel counting shadows in a city, but then using that data to model construction activity. or maybe looking at future up and coming cities by looking at road development and “vascularization”.

We encourage ambitious ideas and as such if you have a really creative idea, but an incomplete execution, we will take that under very strong consideration.

How we will evaluate submissions?

First and foremost, your submission should prove that you can build and develop models. If you opt for option 1 above, you should be able to read relevant research and replicate models for image classification, object localization, and the like. If your idea is really ambitious and won’t show this on its own, maintain a side project that shows that you can perform object detection or semantic segmentation. If you opt for option 2, prove that you can perform cutting edge modeling.

Creative and ambitious ideas: Idea creativity will also go a long way. We are a small team and we are contending against big companies, so out of the box thinking is a prerequisite to joining us. If you can help us generate our next product idea, we will work with you or buy your concept outright, even if you do not win.

Data & Labels

We have not provided much by way of data or labels.

For data, the most accessible and abundant data is the government data (landsat, sentinel, etc). The other data is pretty cool (space net data, dstl) but is limited in that you do not have time-series data, and the data sets are quite small.

For labels, you will again have to be creative. Here, we encourage you to take chances. Can u use hyper spectral data to generate labels, that can then be used to train cheaper, more frequent lower spectral imagery?

Join our slack channel and ask questions or give feedback!