trt_stereo

trt_stereo is an experimental implementation for block matchig stereo algorithm using Cupy, PyTorch and TensorRT. It uses on zero mean sum of absolute difference cost (ZSAD) and compute integer disparity.

Requirements

trt_stereo requires the following libraries:

• cupy
• pillow
• torch
• tensorrt

pip install -r requirement.txt

Usage

# compute disparity and export ONNX model
python Python/zsad_model.py

# build TensorRT Engine
trtexec --shapes=x:1x1x375x450,y:1x62x375x450 --onnx=zsad.onnx --saveEngine=zsad_engine.trt
trtexec --shapes=x:1x1x375x450,y:1x62x375x450 --onnx=zsad.onnx --saveEngine=zsad_engine_fp16.trt --inputIOFormats=fp16:chw --fp16

# execute TensorRT Engine
python Python/zsad_trt.py