Argus: Metric Panoramic 3D Reconstruction for Indoor Scenes

ECCV 2026

Realsee

Quick Start

Clone the repository and install dependencies:

git clone https://github.com/realsee-developer/Argus.git
cd Argus
pip install -r requirements.txt
pip install -e .

Download the pretrained weights:

# Authenticate with HuggingFace (required for gated model)
hf auth login

# Option 1: Auto-download via Python (cached by huggingface_hub)
python -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='RealseeTechnology/argus-realsee3d', filename='argus_realsee3d.pt')"

# Option 2: Manual download
mkdir -p models
hf download RealseeTechnology/argus-realsee3d argus_realsee3d.pt --local-dir models

Run inference with a few lines of code:

import torch
from huggingface_hub import hf_hub_download
from argus.models.argus import Argus
from argus.utils.pose_enc import pose_encoding_to_extri360

# Download model weights (requires: hf auth login)
model_path = hf_hub_download(
    repo_id="RealseeTechnology/argus-realsee3d",
    filename="argus_realsee3d.pt",
)

# Load model
model = Argus(reorder_by_learning_ref=True, restore_metric_scale=True)
model.load_state_dict(torch.load(model_path)["model"], strict=False)
model.eval().cuda()

# Prepare input: panoramic images as tensor [S, 3, H, W], values in [0, 1]
images = ...  # your preprocessed ERP images

with torch.no_grad(), torch.amp.autocast("cuda", dtype=torch.bfloat16):
    predictions = model(images.cuda())

# Extract camera extrinsics
extrinsic, conf = pose_encoding_to_extri360(pose_encoding=predictions["pose_enc"])

# Access depth and 3D points
depth = predictions["depth"]          # [B, S, H, W, 1]
depth_conf = predictions["depth_conf"]  # [B, S, H, W]

Interactive Demo

Launch the Gradio demo for interactive 3D reconstruction and metric measurement:

# Model will be auto-downloaded from HuggingFace if not found locally
# (requires: hf auth login)
python demo_gradio.py

# Or specify a local model path
python demo_gradio.py --model_path models/argus_realsee3d.pt

The demo supports:

• Uploading multiple panoramic images
• Real-time 3D reconstruction with GLB export
• Interactive metric distance measurement between points
• Adjustable confidence thresholds and visualization options

Evaluation

Evaluate on the Realsee3D benchmark:

cd evaluation
python eval.py \
  --model_path ../models/argus_realsee3d.pt \
  --dataset_path /path/to/Realsee3D \
  --split both \
  --ref

Metrics include camera pose accuracy, depth error, point map quality, and covisibility estimation.

Training

Training supports multi-GPU distributed training:

cd training
torchrun --nproc_per_node=8 launch.py --config full

See training/config/full.yaml for the full training configuration.

License

This project is licensed under the Apache License 2.0. The pretrained model weights trained on RealSee3D are released under a non-commercial license, consistent with the RealSee3D dataset license.

Citation

@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}, 
}

Acknowledgements

Argus builds upon VGGT.