arrow_back Supervised-machine-learning without coding
PyTorch Demystified, Why Did I Switch arrow_forward
Object Classification in 3D: Working with LiDAR point clouds
Submitted by Akbar Ladak (@bluplaneter) on Wednesday, 7 June 2017
LiDAR sensors are rapidly becoming mainstream not just to detect obstacles (as in self driving cars), but also to identify object classes for mapping, disaster relief and other use cases.
I will talk about our experience in developing a high accuracy object classification algorithm for LiDAR point clouds that approached human detection using Pattern Recognition over a biased training dataset. I will also touch upon our work using Gain Adversarial Neural Networks (GAAN) to extend the applicability of the algorithm over a wider dataset with minimal effect on identification accuracy.
Introduction [3 min]
- What is LiDAR?
- What are LiDAR point clouds?
- What is the size of typical LiDAR datasets that we encountered?
Why ML is not a magic bullet [5 min]
- Getting our hands around scale
- Data Abstraction & Dimensionality Reduction
- “Knowledge” & Pattern Recognition to the rescue
Extending the algorithm [7 min]
- The need to extend the algorithm
- Introducing Machine Learning
- Using GAAN to train algorithm over diverse datasets
A healthy curiosity / scepticism about Machine Learning & the hype around it.
Akbar Ladak has built technical & business solutions using Computer Vision & Machine Learning at Honeywell & Wipro for over a decade. He is currently the Founder & CEO at Kaaenaat, which builds terrestrial & aerial mapping solutions using images & LiDAR data.