Datasets:

KangLiao
/

IRS-Camera

task_categories:
  - text-to-3d
  - image-to-3d
  - image-text-to-image
  - any-to-any
arxiv: 2609.04196
tags:
  - Camera
  - 3D Vision
  - Spatial AI
  - Physical AI
  - World Model
  - Camera Parameter
  - IRS
  - Indoor
  - Stereo
  - Generation

IRS-Camera

Per-image camera parameter annotations for the IRS dataset (a large synthetic indoor stereo dataset; 188,348 images = the left and right RGB renders of 6 scene archives: Office-1, Office-2, Home-1, Home-2, IRS_small and Store), captioned by the Puffin-World model. More captioned datasets are provided in our Puffin-16M website.

The collage above visualizes the camera maps on sample images — each pair shows the up field (green arrows: the projected gravity-up direction) and the latitude field (colored contours: angle above/below the horizon).

Format

One .tar per shard (irs_Store_0000.tar … ), each containing one .json per image whose name matches the source image stem.

Each JSON holds the predicted monocular camera parameters:

Field	Meaning	Unit
`roll`	camera roll	radians
`pitch`	camera pitch	radians
`vfov`	vertical field-of-view	radians
`k1`	radial distortion coefficient	–
`parse_ok`	whether the model output parsed within valid ranges	bool

Example:

{"roll": 0.0008, "pitch": -0.0564, "vfov": 1.0400, "k1": 0.0000, "parse_ok": true}

Camera Parameter Distributions

Histograms of the predicted roll / pitch / vertical-FoV over the whole dataset (proportion of valid samples per 10° bin; parse_ok=False excluded).

split	roll μ / med / σ	pitch μ / med / σ	FoV μ / med / σ
all (188,254)	0.2° / 0.0° / 4.9°	−8.3° / −5.0° / 14.8°	54.8° / 56.7° / 6.1°

Roll is tightly peaked at 0° (96% of images within ±5°) — the virtual stereo rig is kept level in almost every rendered trajectory.
Pitch is clearly negative (μ ≈ −8.3°, σ ≈ 14.8°): indoor robot-height viewpoints look downward far more often than upward.
FoV is narrow and highly concentrated (σ ≈ 6.1°, 75% within 50–60°), as expected from a synthetic dataset rendered with a near-fixed virtual lens.

If you'd like a dataset with a more diverse and uniform distribution of camera parameters, please refer to our Puffin-4M and Puffin-16M datasets.

Dataset Download

You can download the entire dataset using the following command:

hf download KangLiao/IRS-Camera --repo-type dataset

From Camera Parameters to Up and Latitude Fields

The released (roll, pitch, vfov, k1) annotations can be converted into the dense perspective-field representation used by Puffin-World. The conversion computes focal length from vfov, constructs a radial camera, and maps roll and pitch to the gravity direction. get_perspective_field then returns a normalized 2-channel up field and a 1-channel latitude field. The up field encodes the projected world-up direction at every pixel, while the latitude field measures each viewing ray's angular elevation relative to the horizon.

import json
import torch
from scripts.camera.geometry.camera import SimpleRadial
from scripts.camera.geometry.gravity import Gravity
from scripts.camera.geometry.perspective_fields import get_perspective_field
from scripts.camera.utils.conversions import fov2focal

with open("camera.json") as f:
    annotation = json.load(f)
roll, pitch, vfov, k1 = (
    annotation[key] for key in ("roll", "pitch", "vfov", "k1")
)

H, W = 512, 512
f = float(fov2focal(torch.tensor(vfov), H))
camera = SimpleRadial(torch.tensor(
    [W, H, f, f, W / 2, H / 2, k1, 0.0]
).float()).scale(torch.tensor([1.0, 1.0]))
gravity = Gravity.from_rp(torch.tensor(roll), torch.tensor(pitch))
up_field, latitude_field = get_perspective_field(camera, gravity)
# Shapes: [1, 2, H, W] and [1, 1, H, W]

Run the example from the Puffin-World directory. Use plot_vector_fields and plot_latitudes to render up-field arrows and latitude heatmaps or contours. See save_pf_visualization for an end-to-end visualization example.

Caption Pipeline

Beyond this captioned dataset, we also release a complete captioning pipeline for annotating camera parameters for arbitrary datasets, analyzing camera parameter distributions, and visualizing the corresponding camera maps. The pipeline is available in our GitHub repository.

Citation

If you find the captioned dataset useful for your research or applications, please cite the following papers using these BibTeX entries:

  @article{liao2025puffin,
    title={Thinking with Camera: A Unified Multimodal Model for Camera-Centric Understanding and Generation},
    author={Liao, Kang and Wu, Size and Wu, Zhonghua and Jin, Linyi and Wang, Chao and Wang, Yikai and Wang, Fei and Li, Wei and Loy, Chen Change},
    journal={arXiv preprint arXiv:2510.08673},
    year={2025}
  }

  @article{liao2026puffinworld,
    title={Puffin-World: Scaling a Unified Multimodal Model with Native 3D World States},
    author={Liao, Kang and Luo, Yihang and Wu, Xiao-Ming and Jin, Linyi and Wu, Size and Lin, Chunyu and Zhao, Yao and Wang, Fei and Li, Wei and Loy, Chen Change},
    journal={arXiv preprint arXiv:2609.04196},
    year={2026}
  }