DavidSeyserHF/resireg-mini-onnx

ReSiReg Mini - Compiled ONNX for OAK

This repository contains a compiled ONNX export of the ReSiReg Mini image encoder, prepared for OpenVINO 2021.4 and OAK / DepthAI deployment.

Important: this is not the full ReSiReg model. The text encoder is not included in this ONNX export.

ReSiReg is a vision-language model for spatially consistent language grounding in robotic tasks. This export keeps the dense image branch only: it turns a normalized 512x512 RGB image tensor into dense ReSiReg features that can be compared with text or prompt embeddings on the host.

The graph is static and deployment-oriented:

• input: pixel_values with shape [1, 3, 512, 512]
• output: dense image features with shape [1, 512, 32, 32]
• target runtime: OpenVINO 2021.4 / MyriadX OAK devices

Text encoding must run outside this compiled model. Generate or load prompt embeddings on the host, then compare them against the dense image features returned by the ONNX model.

Export Notes

The model was exported from SimonSchwaiger/resireg_mini and rewritten for OpenVINO 2021.4 compatibility. The final ONNX output was checked against the local PyTorch model on a deterministic test input:

np.allclose(rtol=1e-4, atol=1e-4): True

Citation

@article{schwaiger2026_resireg,
  title   = {{ReSiReg}: Towards Spatially Consistent Semantics in Language-Conditioned Robotic Tasks},
  author  = {Schwaiger, Simon and Seyser, David and Scherl, Alessandro and W{\"o}ber, Wilfried and Steinbauer-Wagner, Gerald},
  journal = {arXiv preprint arXiv:2606.19088},
  year    = {2026},
  url     = {https://arxiv.org/abs/2606.19088}
}

License

This export follows the license and upstream restrictions of the original ReSiReg Mini model and its dependencies. In particular, the upstream EUPE component is noncommercial-research licensed, so downstream use must comply with that restriction.