M2H-MX Multi-Task Weights

This repository hosts model-only weights for M2H-MX: Multi-Task Semantic and Geometric Perception for Real-Time Monocular 3D Scene Graph Construction.

M2H-MX is a multi-task dense visual perception model, not a semantic-segmentation-only model. Given a monocular RGB image, the network can predict:

• metric depth or disparity, depending on the dataset configuration;
• semantic segmentation logits;
• surface normals;
• edge maps.

Depth and semantics are the primary deployment outputs used by Mono-Hydra++ or a compatible mapping backend for metric-semantic mapping and downstream 3D scene graph construction. Surface normals and edges are auxiliary training heads used to improve geometric and semantic consistency. The network improves the dense evidence used by the mapping backend; it does not directly predict the 3D scene graph.

Code and instructions: https://github.com/BavanthaU/m2h_mx

Artifacts

Dataset	Variant	File	Paper result
NYUDv2	M2H-MX-L	weights/nyudv2/m2h_mx_l_nyudv2_weights.pt	mIoU 65.60, depth RMSE 0.3800
NYUDv2	M2H-MX-B	weights/nyudv2/m2h_mx_b_nyudv2_weights.pt	mIoU 61.80, depth RMSE 0.4170
ScanNet	M2H-MX-L	weights/scannet/m2h_mx_l_scannet_weights.pt	ScanNet25k mIoU 76.10, depth RMSE 0.2210; Mono-Hydra++ ATE 6.91 cm
ScanNet	M2H-MX-B	weights/scannet/m2h_mx_b_scannet_weights.pt	Base variant artifact
Cityscapes	M2H-MX-L	weights/cityscapes/m2h_mx_l_cityscapes_weights.pt	mIoU 82.28, disparity RMSE 3.89

These are model-only state dictionaries. They do not include optimizer, scheduler, gradient scaler, or EMA state.

Download

From the code repository:

python3 scripts/download_weights.py --repo-id Bavantha11/m2h-mx --verify

Citation

@misc{udugama2026m2hmxmultitaskdensevisual,
  title={M2H-MX: Multi-Task Semantic and Geometric Perception for Real-Time Monocular 3D Scene Graph Construction},
  author={U. V. B. L. Udugama and George Vosselman and Francesco Nex},
  year={2026},
  eprint={2603.29236},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.29236},
}

@misc{udugama2026monohydrarealtimemonocularscene,
  title={Mono-Hydra++: Real-Time Monocular Scene Graph Construction with Multi-Task Learning for 3D Indoor Mapping},
  author={U. V. B. L. Udugama and George Vosselman and Francesco Nex},
  year={2026},
  eprint={2605.17661},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2605.17661},
}