Project Brief
Created scripts for a C++ SDK to deploy a defect segmentation model for real-time usage and compare performance/speed trade-off between different precisions
Accomplishments
- Implemented a script to double confirm performance (mIOU) of TensorRT engine with reference to original model (from pth) performance
- Added support for FP32/FP16/INT8 precision conversion from onnx to TensorRT engine
- Compared performance/speed trade-off between different precisions and modes of deployment
Gallery
Visual confirmation with reference to original material with defect and comparison between original and TensorRT models
Table for performance/speed trade-off between different precisions and modes of deployment (not all experiments included due to confidentiality)