Geometric deep learning and TensorRT at NVIDIA GTC 2022

About the event

GTC is NVIDIA's conference on AI and accelerated computing. GTC 2022 ran online from 21 to 24 March 2022.

The session

“Save Thousands of Man-hours with Geometric Deep Learning and TensorRT” is session S41247 of GTC 2022. The speakers were Mayur Kulkarni and Sujit Nalawade. It ran on Thursday 24 March 2022, from 4:30 to 5:20 pm IST. NVIDIA rates the session “Technical, Advanced” and files it under deep learning inference.

The subject is lidar point clouds of power lines. Utility networks run electric poles for thousands of kilometres. One tree that grows too close to a high-voltage line can cause an outage or a wildfire. To find such trees by hand is slow and costly.

The session covers two problems:

• Segmentation: classify point clouds of more than 100 million points into classes such as electric poles and buildings. The tools are PyTorch Geometric, Torch Points 3D and PyTorch Cluster.
• Inference at scale: dense point clouds need a large amount of compute. The team used CUDA and NVIDIA TensorRT to scale the inference pipeline on NVIDIA A100 servers.

Why it matters for roads

The power-line project was not about traffic, but the skills carried over. Today every model in the Recon Global Shutter Camera runs through TensorRT on an NVIDIA Jetson module inside the camera. In 2026 our poster on Recon was shown at NVIDIA GTC 2026, and our research team won Track 3 of the AI City Challenge.

It was our second GTC session, after object detection at 1770 FPS in 2021. Our team has built edge AI for traffic and vehicle recognition since 2019. Read more on our research page.