panodac 0.1.1

Plug-and-play metric depth estimation for any camera (perspective, fisheye, 360°). Based on Depth Any Camera (CVPR 2025).

panodac

Metric depth estimation for any camera. Perspective, fisheye, 360° panorama.

Based on Depth Any Camera (CVPR 2025).

Installation

pip install panodac

Usage

import panodac

depth = panodac.predict("photo.jpg")
# depth is a numpy array (H, W) with metric depth in meters

Panorama Seam Blending

ERP panoramas wrap horizontally, but CNN padding can introduce a visible seam at the left/right boundary. By default, panodac applies a Poisson-based seam correction when a panorama is detected.

import panodac

# Default: seam correction enabled for panoramas
depth = panodac.predict("panorama.jpg")

# Disable seam correction if you want raw output
depth_raw = panodac.predict("panorama.jpg", fix_panorama_seam=False)

Models

Model	Use Case	Speed	Quality
outdoor-resnet101	Outdoor	Fast	Good
outdoor-swinl	Outdoor	Slow	Best
indoor-resnet101	Indoor	Fast	Good
indoor-swinl	Indoor	Slow	Best

Models auto-download from HuggingFace on first use (~500MB each).

# Use a specific model
depth = panodac.predict("panorama.jpg", model="outdoor-swinl")

# List available models
print(panodac.list_models())
# ['outdoor-resnet101', 'outdoor-swinl', 'indoor-resnet101', 'indoor-swinl']

Documentation

See yz3440.github.io/panodac for full API reference and examples.

Credits

Based on Depth Any Camera (DAC) by Yuliang Guo et al.

• Paper
• Original Repository
• Project Page

@article{guo2025depthany,
  title={Depth Any Camera: Zero-Shot Metric Depth Estimation from Any Camera},
  author={Guo, Yuliang and Garg, Sparsh and Ren, Xuan and ElSayed, Mohamed and Guizilini, Vitor},
  journal={CVPR},
  year={2025},
}

License

MIT