Abstract: 

Scalable feature detection and learning is very important in the autonomous world. One of great sources of these features is lidar data (point cloud in 3D space) that we collect from vehicles. However, this data is not easy to deal with given its irregular pattern. There has been several works on lidar based feature detection using traditional approaches that deeply relies on 3D geometry. However, we believe this approach is not very generalized and for each type of feature we need separate algorithms which makes lidar data processing less effective from feature detection's perspective. In this talk we introduce few deep learning based approaches (MaskRCNN etc) to show how we can exploit them to mitigate this problem.

Bio: 

Shubhabrata is a senior data scientist in HERE Technologies. He is actively involved in Computer Vision and "UnderstandingTraffic Movement" projects for the organisation. He has vast experience in working with different domains including text mining, NLP, sentiment analysis. Previously he worked at AI startup RealEyes on facial recognition for media. Shubhabrata finished his PhD on Signal Processing application from École Normale Supérieure de Lyon in France with a summa cum laude.