Solving SpaceNet Road Detection Challenge With Deep Learning

Originally published at: Solving SpaceNet Road Detection Challenge With Deep Learning | NVIDIA Technical Blog

It’s that time again — SpaceNet raised the bar in their third challenge to detect road-networks in overhead imagery around the world. Today, map features such as roads, building footprints, and points of interest are primarily created through manual techniques. In the third SpaceNet challenge, competitors were tasked with finding automated methods for extracting map-ready…

Hi guys, thank you for a nice article but I would like to point out one thing regarding the floodfill.
Actually you do not have to do this. The hosts of the competitions provided road width in their geojson files.
I tried it creating masks from this - and succeeded - loss was lower by ~25%.
But graph creation algorithm worked worse.

See my post for some details - https://spark-in.me/post/sp... - just search for the word "wide"

Finally, we take creative liberties to think about how we might apply these types of deep learning solutions in a broader operational sense using conditional random fields, percolation theory, and reinforcement learning.

If you use the same tiramisu architecture for two experiments, one using an 8-channel input, the other using a 3-channel input, will your GPU footprint be larger on the 8-channel experiment? Is there a downside in training performance by adding many more channels? I have tested 1-channel and 3-channel tiramisu and have found the GPU footprint and training performance to be very similar.

https://hardwarereview79405...
https://www.blogger.com/u/1...
https://wordpress.com/block...
https://www.blogger.com/u/3...
https://greenwheelbarrow.we...
https://videoshopgalley.wee...
https://wheelbarrow4201.wix...
https://sitebuilder.yola.co...
https://sites.google.com/s/...
https://sites.google.com/s/...