[Pycon] [new paper] "Shubham Goel" - Building footprint extraction using semantic segmentation

[info a pycon.it]

Title: Building footprint extraction using semantic segmentation
Duration: 45 (includes Q&A)
Q&A Session: 15
Language: en
Type: Talk

Abstract: The talk covers how Convolution Neural Networks can be used to extract building polygons from high resolution satellite imagery. The complete pipeline, from pre-processing the data for making it suitable for model ingestion, to tuning the model architecture and hyperparameters, and eventually post-processing the model predictions to generate the final polygons, is discussed in detail.

The talk would start by introducing the importance of remote sensing and how solutions built on top can scale massively. It would then move to the specific problem in hand, which is to extract building footprint. The next step would be giving a high level overview of setting up a deep learning environment suitable for training. 

The major portion of the talk would discuss the model pipeline, which would include gathering data from varied geographies, converting and sampling it for our use case, trying out several CNN architectures alongside varying hyperparameters, and post-processing the final result into geo-referenced polygons which can be overlaid on Google maps. It would also consist of a portion which discusses the importance of using the current loss function (Dice loss in this case) and the scale of impact it can have on the results.

The talk would conclude by citing several other use cases which can be built using a similar pipeline.

Tags: [u'deep learning', u'Satellite', u'computer-vision']