SEAR: Simple and Efficient Adaptation of Visual Geometric Transformers for Unpaired RGB+Thermal 3D Reconstruction (ECCV 2026)
Try the live demo β SEAR demo Space β upload an RGB set and a thermal set, get one aligned 3D scene + camera poses, no setup.
SEAR adapts feed-forward visual geometry models (VGGT) to mixed RGB-thermal input. Give it an RGB image set and a thermal image set of the same scene, captured along different trajectories with no shared camera poses. In a single feed-forward pass it recovers one geometrically consistent 3D scene together with the camera poses of both modalities. It outperforms state-of-the-art 3D reconstruction and pose estimation methods on RGB-T data (e.g. over 29% in AUC@30) with negligible inference overhead over the original pretrained model, and stays reliable in low light and dense smoke.
Resources
- arXiv paper Β· HF papers page
- Project page β qualitative results, interactive comparisons, ablations
- Demo Space β interactive reconstruction, also exposes an API/
reconstructendpoint - Reference code
- ThermoScenes dataset β RGB and thermal sequences at different times, viewpoints, illumination
Usage
This repo provides the SEAR checkpoints (SEAR_X.ckpt) built on facebook/VGGT-1B.
The quickest path is the self-contained loader shipped in the demo Space:
from sear_model import load_sear_model # from the demo Space repo
model = load_sear_model() # downloads facebook/VGGT-1B + this checkpoint
# model(images, thermal_mask) # images: [B, S, 3, H, W]; thermal_mask: [B, S] bool
Citation
SEAR was published at ECCV 2026. If you use SEAR or ThermoScenes, please cite the proceedings version:
@inproceedings{skorokhodov2026sear,
title={SEAR: Simple and Efficient Adaptation of Visual Geometric Transformers for Unpaired R GB-Thermal 3D Reconstruction},
author={Skorokhodov, Vsevolod and Xu, Chenghao and Sun, Shuo and Fink, Olga and Mielle, Malcolm},
booktitle={European Conference on Computer Vision},
pages={542--560},
year={2026},
organization={Springer}
}
Model tree for MalcolmMielle/SEAR
Base model
facebook/VGGT-1B


