Many Eyes, One World: Feed-Forward 3D Reconstruction from Mixed Cameras
Paper • 2609.35658 • Published
Request access to MEOW weights
Email [email protected] with your Hugging Face username, affiliation, and intended use, then submit your request here. Access is reviewed manually.
Log in or Sign Up to review the conditions and access this model content.
MEOW: Many Eyes, One World
Feed-Forward 3D Reconstruction from Mixed Cameras
Code and documentation · Paper · Laser-scan benchmark
MEOW jointly reconstructs metric pointmaps and camera poses from mixed perspective, fisheye, and full 360-degree panoramic images in one forward pass. No camera calibration, distortion parameters, camera-type labels, or poses are supplied to the model. Full panoramas are identified from their pixels for panorama-aware processing.
Checkpoint
| File | Model | Format | Size |
|---|---|---|---|
MEOW-sB25.pth |
Final interpolated MEOW model (sB25) | PyTorch checkpoint, FP32 | 5.05 GB |
The checkpoint contains the complete model parameters, including the aspect-ratio encoder. It is not quantised and does not include optimiser state. Use the :wrap model variant in the MEOW evaluation scripts; do not load it as the unmodified MapAnything baseline.
Access
Download access is reviewed manually, not granted automatically. Before approval, contact the authors at [email protected] with your Hugging Face username, affiliation, and intended use, and submit an access request from the same Hugging Face account.
Download and use
After approval:
pip install -U huggingface_hub
hf auth login
hf download qgli/MEOW-4D-Space-Reconstruction MEOW-sB25.pth --local-dir checkpoints
Clone the code repository and follow the installation instructions. From the repository root, for a mixed-camera tuple set:
export CKPT="$PWD/checkpoints/MEOW-sB25.pth"
python scripts/realset/predict_ours.py \
--tuples /path/to/tuples.json --frames-root /path/to/frames \
--ckpt "MEOW=$CKPT:wrap" --input-resize squeeze --pano auto --out preds
The full-field-of-view resizing and image-only panorama detector are selected explicitly above. See evaluation documentation for the tuple format, outputs, and dataset-specific settings.
Citation
@article{li2026meow,
title = {Many Eyes, One World: Feed-Forward 3D Reconstruction from Mixed Cameras},
author = {Li, Qiaoge and Zhan, Yifan and Yang, Haijun and Liu, Haiyang and Cai, Yiyi and Luo, Chenchi},
journal = {arXiv preprint arXiv:2609.35658},
year = {2026}
}
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support