argus-realsee3d / README.md
lilinyuan001
feat: add demo address
5fcd966
|
Raw
History Blame Contribute Delete
4.82 kB
metadata
license: cc-by-nc-4.0
library_name: pytorch
pipeline_tag: image-to-3d
tags:
  - 3d-reconstruction
  - panoramic
  - depth-estimation
  - camera-pose-estimation
  - point-cloud
  - indoor-scenes
  - feed-forward
gated: true
extra_gated_heading: Request access to Argus
extra_gated_description: >-
  Argus is released for non-commercial research and educational use only. We
  review access requests and may take 2-3 business days to respond.
extra_gated_prompt: >-
  By requesting access you agree to use Argus and the associated weights for
  non-commercial purposes only, in accordance with the CC BY-NC 4.0 license.
extra_gated_fields:
  Full name: text
  Affiliation / Organization: text
  Country: country
  Intended use:
    type: select
    options:
      - Academic research
      - Education
      - Personal / non-commercial project
      - label: Other
        value: other
  Brief description of intended use: text
  I confirm I will use Argus for non-commercial purposes only: checkbox
  I agree to cite the Argus paper in any resulting work: checkbox
extra_gated_button_content: Agree and request access

Argus: Metric Panoramic 3D Reconstruction for Indoor Scenes

ECCV 2026

Project Page arXiv HuggingFace Model HuggingFace Demo RealSee3D Dataset

Realsee

Argus is a feed-forward network for metric panoramic 3D reconstruction of indoor scenes. Given sparse, unordered panoramic captures, it jointly predicts camera pose, metric depth, and point cloud reconstruction in a shared metric world frame.

This repository hosts the pretrained Argus weights (argus_realsee3d.pt).

Overview

Metric feed-forward 3D reconstruction for panoramic data has remained under-explored due to the lack of large-scale panoramic RGB-D training data. This work introduces:

  • Realsee3D — a hybrid dataset of 10K indoor scenes (1K real, 9K synthetic) with 299K panoramic viewpoints and precise metric annotations.
  • Argus — a feed-forward network trained on Realsee3D for metric panoramic 3D reconstruction.

In the sparse unordered capture setting of Realsee3D, a poorly chosen coordinate anchor can cause global pose drift. Argus addresses this with a learned covisibility module that selects the geometrically optimal reference view to anchor the metric world frame. To improve multi-task learning, the bidirectional pixel-to-world mapping is decomposed into interpretable sub-steps with per-step supervision and cross-coordinate joint constraints, reinforcing geometric consistency across prediction branches.

On the Realsee3D benchmark, Argus achieves state-of-the-art metric performance in camera pose estimation, depth estimation, and point cloud reconstruction.

Files

File Description
argus_realsee3d.pt Pretrained Argus model weights (PyTorch checkpoint)

Usage

For the model architecture, inference pipeline, and example code, see the official repository: https://github.com/realsee-developer/Argus

Citation

If you use Argus or the Realsee3D dataset, please cite:

@misc{li2026argusmetricpanoramic3d,
      title={Argus: Metric Panoramic 3D Reconstruction for Indoor Scenes}, 
      author={Xi Li and Linyuan Li and Yan Wu and Tong Rao and Kai Zhang and Xinchen Hui and Cihui Pan},
      year={2026},
      eprint={2606.30047},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.30047}, 
}