SuperPoint TensorRT Optimization

Optimizing SuperPoint keypoint detector for edge deployment using TensorRT. PyTorch → ONNX → TensorRT FP16, targeting NVIDIA Jetson Orin.

Pipeline

1. Export SuperPoint PyTorch model to ONNX (export_onnx.py)
2. Simplify ONNX graph with onnx-simplifier
3. Compile TensorRT FP32 and FP16 engines (build_engines.py)
4. Benchmark inference latency (benchmark.py)

Results

Hardware: NVIDIA Tesla T4, CUDA 13.0, TensorRT 10.11

Backend	Mean Latency	P99 Latency	FPS
PyTorch CUDA	14.66 ms	15.28 ms	68
TensorRT FP32	0.61 ms	1.87 ms	1642
TensorRT FP16	0.45 ms	1.42 ms	2203

TensorRT FP16 achieves 32× speedup over PyTorch baseline.

Setup

pip install torch onnx onnx-simplifier tensorrt pycuda
python export_onnx.py        # generates superpoint_simplified.onnx
python build_engines.py      # generates fp32 and fp16 .engine files
python benchmark.py          # runs latency benchmark