LiDAR Perception Researcher at NVIDIA

Conducting research towards improving 3D object detection performance.

I worked on the LiDAR perception sub-team of the autonomous vehicles organization. LiDAR is an awesome sensor modality, which is similar to computer vision but with a couple of very important differences. In my experiments to improve object detection performance, I was doing a lot of 3D point cloud processing, with dense and sparse 3D convolutional networks.

Going Sparse

LiDAR data is very sparse, so one of my big contributions was getting sparse tensors and sparse tensor networks working throughout a whole detection pipeline. Sparse tensors allowed me to reduce the memory footprint of my network by 98%, while improving F-scores. I was using Minkowski Engine, an open source sparse tensor deep learning library from NVIDIA, which has great integration with PyTorch.

What I Learned

• Evaluating modern point cloud object detection methods on real data.
• Staying up to date on literature in the field.
• Working in a highly structured/safety-certified development environment.
• Production machine learning engineering, with changing requirements, compute resources, and approaches.