LightGlue-ONNX: Local Feature Matching at Light Speed

• URL: https://github.com/fabio-sim/LightGlue-ONNX
• Date Saved: 2026-09-12
• Source: GitHub
• Tags: ai-tools, dev-tools
• Repo: https://github.com/fabio-sim/LightGlue-ONNX (776 stars, Apache-2.0)

Summary

ONNX-compatible export of LightGlue (local feature matching) — removes PyTorch dependency for production deployment.

• Exports full extractor+matcher pipeline to ONNX for interop across platforms
• Supports multiple backends: TensorRT, OpenVINO, CUDA, CPU
• Extractors: SuperPoint, DISK, RaCo-ALIKED
• Latest (Jul 2026): optimized RaCo-ALIKED-LightGlue+ with TensorRT — significant speedup over torch.compile()
• FP8 quantization workflow (NVIDIA ModelOpt Q/DQ export) for even faster TensorRT inference
• CLI tool via typer: lightglue-onnx export/infer/trtexec
• Dynamic batch size support, FlashAttention-2 via onnxruntime >= 1.16
• TopK-trick optimization: ~30% speedup by eliminating ArgMax
• Browser WebGPU demo included
• Uses uv for dependency management with granular install groups (cpu, cuda, trt, export)
• Integrated into Kornia (kornia.feature.OnnxLightGlue)

Useful for deploying feature matching in production pipelines (SLAM, visual localization, homography estimation) without Python/PyTorch overhead. Pairs well with the ABot-Recon style 3D reconstruction workflows.