agrippa-stereocam

IGEV++ Setup

This guide covers the current IGEV++ export and runtime workflow for ag-cam-tools.

Two phases

1. Export the PyTorch checkpoint to ONNX on a development machine
2. Run the exported ONNX model through the in-process ONNX Runtime backend on the target machine

The export is tied to the calibration resolution and disparity range.

Phase 1: export to ONNX

Clone the IGEV++ repo

cd ~/code
git clone https://github.com/gangweiX/IGEV-plusplus.git

Download a checkpoint

Use the SceneFlow checkpoint from the upstream project and place it under:

~/code/IGEV-plusplus/pretrained_models/igev_plusplus/

Create the export environment

cd backends
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements-onnx-export.txt
python -m ipykernel install --user --name agrippa-export --display-name "agrippa-export"

Run the notebook

Open backends/1.Export_IGEV.ipynb and run it with the agrippa-export kernel.

The notebook:

• reads calibration metadata,
• determines the export resolution,
• exports ONNX at a fixed shape,
• simplifies and validates the result.

If the calibration changes materially, re-export the model.

Phase 2: runtime inference

Install ONNX Runtime

The runtime build can discover ONNX Runtime either through pkg-config or ONNXRUNTIME_HOME.

export ONNXRUNTIME_HOME=/path/to/onnxruntime

Rebuild the tool

make clean && make

The build should indicate that ONNX Runtime support is enabled.

Run the model

ag-cam-tools depth-preview-neural \
    -a 192.168.0.201 \
    -A \
    -r calibration/calibration_YYYYMMDD_HHMMSS \
    --stereo-backend onnx \
    --model-path models/igev_plusplus.onnx

Troubleshooting highlights

• If the notebook cannot import IGEV modules, check the configured source path.
• If export fails with CUDA memory issues, force CPU export.
• If inference is slow on Jetson, confirm the CUDA execution provider is actually active.
• If the ONNX model resolution differs from the runtime image size, re-export to match the calibrated setup.

This site is open source. Improve this page.