Code for “EvidMTL: Evidential Multi-Task Learning for Uncertainty-Aware Semantic Surface Mapping from Monocular RGB Images”, by Rohit Menon, Nils Dengler, Sicong Pan, Gokul Krishna Chenchani, and Maren Bennewitz.
Given a monocular RGB image, EvidMTL jointly predicts semantic labels and depth together with their uncertainties. EvidKimera uses those uncertainty estimates to produce an uncertainty-aware semantic TSDF map.
This repository is the entry point for the EvidMTL and EvidKimera uncertainty-aware semantic surface mapping pipeline. It pins the three public components as Git submodules:
| Submodule | Purpose | Upstream foundation | License |
|---|---|---|---|
evidmtl |
Evidential semantic segmentation and monocular depth estimation | SwinMTL | GPL-3.0 |
evidkimera |
Uncertainty-aware semantic TSDF fusion | Kimera-Semantics | BSD-2-Clause |
voxblox |
Custom Voxblox and voxblox_ros support for evidential fusion |
Voxblox | BSD-3-Clause |
The 39-second project video is kept as MP4 to preserve its 1080p quality while remaining smaller than a full-length GIF. Click the preview to play it.
The project introduced:
- EvidMTL: evidential heads for joint semantic segmentation and monocular depth, semantic and depth uncertainty outputs, and an evidential SiLog depth objective with uncertainty regularization and a Normal-Inverse-Gamma prior;
- EvidKimera: uncertainty-aware depth weighting, evidence-aware semantic fusion, voxel-level uncertainty, and viewpoint-similarity weighting for semantic TSDF mapping; and
- Voxblox integration: extended point, voxel, mesh-message, ROS conversion, and integrator interfaces that carry semantic evidence and depth uncertainty through the mapping pipeline.
These are extensions of SwinMTL, Kimera-Semantics, and Voxblox respectively; the original projects and publications are credited below and in each component repository.
EvidMTL uses a shared encoder-decoder with evidential semantic and depth heads. The resulting uncertainty estimates are evaluated against the corresponding prediction errors.
EvidKimera integrates the evidential predictions into a semantic TSDF. Depth uncertainty controls geometric fusion weights, while semantic evidence updates the voxel semantic state and its uncertainty.
The paper evaluates zero-shot depth uncertainty on ScanNetV2 after training on NYUDepthV2. The visualization below compares depth error, aleatoric uncertainty, and epistemic uncertainty for EvidMTL and two ablations.
Figures and descriptions in this section are taken from the arXiv paper.
git clone --recurse-submodules https://github.com/HumanoidsBonn/evidential_mapping.gitFor an existing clone:
git submodule update --init --recursiveThe EvidMTL release uses two checkpoints:
silog_klmuprior_alpha1.01_gamma0.001_betacalc_final_model.ckpt: noisy ground-truth mean prior;silog_klmuprior_alpha2_beta0.1_gamma1.0_no_noise_final_model.ckpt: ground-truth mean prior without added noise.
They are distributed separately through the EvidMTL Hugging Face model repository.
If you use this software, please cite our EvidMTL paper and the three foundational papers for SwinMTL, Kimera, and Voxblox:
@inproceedings{menon2025evidmtl,
author={Menon, Rohit and Dengler, Nils and Pan, Sicong and Chenchani, Gokul Krishna and Bennewitz, Maren},
title={EvidMTL: Evidential Multi-Task Learning for Uncertainty-Aware Semantic Surface Mapping from Monocular RGB Images},
booktitle={2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year={2025},
doi={10.1109/IROS60139.2025.11247528}
}
@inproceedings{taghavi2024swinmtl,
author={Taghavi, Pardis and Langari, Reza and Pandey, Gaurav},
title={SwinMTL: A Shared Architecture for Simultaneous Depth Estimation and Semantic Segmentation from Monocular Camera Images},
booktitle={2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year={2024},
url={https://arxiv.org/abs/2403.10662}
}
@inproceedings{rosinol2020kimera,
author={Rosinol, Antoni and Abate, Marcus and Chang, Yun and Carlone, Luca},
title={Kimera: an Open-Source Library for Real-Time Metric-Semantic Localization and Mapping},
booktitle={2020 IEEE International Conference on Robotics and Automation (ICRA)},
year={2020},
doi={10.1109/ICRA40945.2020.9196885}
}
@inproceedings{oleynikova2017voxblox,
author={Oleynikova, Helen and Taylor, Zachary and Fehr, Marius and Nieto, Juan and Siegwart, Roland},
title={Voxblox: Incremental 3D Euclidean Signed Distance Fields for On-Board MAV Planning},
booktitle={2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
year={2017},
doi={10.1109/IROS.2017.8202315}
}The EvidMTL paper is available from IEEE Xplore and arXiv.
Each submodule retains its own license. See evidmtl/LICENSE,
evidkimera/LICENSE.BSD, and voxblox/LICENSE. The umbrella repository
does not relicense the component code.



