SURE-Map: Self-Correcting Streaming Geometric Foundation Models

Mingkai Liu, Hao Zhao, Xingxing Zuo

Paper · Code · Project page · Video

SURE-Map equips streaming geometric foundation models with cross-view geometric uncertainty and multi-timescale self-correction. Cross-view uncertainty assesses the geometric consistency of jointly predicted pose and depth, supporting dense-point filtering and local translation optimization. Sparse keyframe-window inference provides longer-range geometric evidence for scale recalibration.

Checkpoint

uncertainty.pt is the trained cross-view geometric uncertainty head checkpoint used in the reported experiments. It is used together with the SURE-Map implementation and a separately downloaded LingBot-Map backbone; it is not a standalone reconstruction model.

The checkpoint is copied without modification from the official SURE-Map repository.

• File size: 199,159,826 bytes.
• SHA-256: a1b8cdd76ae33768e738554cece65b0113770e48824064c7a92b59752b06e1a7.
• Training data: TartanAir v1; the backbone is frozen during uncertainty-head training.

Download and use

Clone the code and follow its environment setup instructions:

git clone https://github.com/RCL-Robotics/SURE-map.git SURE-Map
cd SURE-Map
mkdir -p checkpoints

Download the uncertainty checkpoint into checkpoints/uncertainty.pt:

curl -L --fail \
  https://huggingface.co/milchstrasse/SURE-Map/resolve/main/uncertainty.pt \
  -o checkpoints/uncertainty.pt

Download the backbone separately:

curl -L --fail \
  https://huggingface.co/robbyant/lingbot-map/resolve/main/lingbot-map.pt \
  -o checkpoints/lingbot.pt

The SURE-Map inference code constructs FlowSigmaHead and loads the checkpoint's sigma_head state dictionary. Use the repository's inference and evaluation entry points rather than a Transformers from_pretrained call.

For example, after preparing KITTI and setting dataset.root in online/configs/kitti.yaml:

python online/run_kitti.py --config online/configs/kitti.yaml

See the code repository for Oxford Spires, VBR, Neural RGB-D, 7-Scenes, and DTU preparation and evaluation instructions. This checkpoint is provided for the documented LingBot-Map-based configuration; compatibility with other backbones is not established.

Citation

@misc{liu2026suremapselfcorrectingstreaminggeometric,
  title         = {SURE-Map: Self-Correcting Streaming Geometric Foundation Models},
  author        = {Mingkai Liu and Hao Zhao and Xingxing Zuo},
  year          = {2026},
  eprint        = {2609.15795},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2609.15795}
}

Licensing

The upstream repository licenses its original source code under Apache-2.0 and preserves separate licenses for third-party components. It does not explicitly state a separate license for this checkpoint. See the upstream license information.