From Proof of Concept to Industrial Products by Leveraging TensorRT and GPUs
, CTO, Bilberry
Sorting and spot-spraying machines in agriculture need to be precise and fast in order to deliver enough capacity to be commercially viable. These computer vision-based products require a processing capability that can trigger an action in less than 100ms. By leveraging TensorRT, we validated our value proposition on proof of concept based on a Geforce GPU. Once validated, we seamlessly migrated to a Jetson TX2 using Float16 precision and model pruning optimizations to improve inference speed. Finally, we were able to rationalize our hardware even further by leveraging the full power of an AGX Xavier, which enabled us to process multiple cameras concurrently on one processing unit. This final step was enabled by TensorRT INT8 inference optimization. We'll detail how you can quickly validate your idea, limit the time needed to optimize performance, and how TensorRT can help you quickly commercialize it.