Object detection at 1770 FPS at NVIDIA GTC 2021

About the event

GTC is NVIDIA's conference on AI and accelerated computing. GTC 2021 ran online from 12 to 16 April 2021.

The session

“Object Detection at 1770 FPS: Designing CNNs that Scale Using DeepStream SDK” is session S31215 of GTC 2021, a recorded talk of 40 minutes. The speakers were Mayur Kulkarni and Sujit Nalawade. NVIDIA rates the session “Technical, Intermediate” and files it under computer vision and image processing.

The session answers three questions about CNNs on NVIDIA GPUs:

• Design: what makes current object detection models efficient, and how to design a model that scales.
• Benchmark: how to measure a model and its pipeline with tools such as NVTX.
• Deploy: how to put the models into production with the NVIDIA DeepStream SDK.

It also covers what the team learned from video analytics pipelines in production, from research to deployment at scale.

Why it matters for roads

A traffic camera must read every vehicle in every frame, on a small computer inside the camera. The same work on model design, measurement and deployment is behind the Recon Global Shutter Camera, where several models run at the same time on an NVIDIA Jetson module.

The team returned to GTC with a session in 2022 and a poster in 2026. Read more on our research page.