README.md

metadata

license: apache-2.0
task_categories:
  - image-feature-extraction
tags:
  - camera
  - camera-calibration
  - imagenet
  - camera-parameters
size_categories:
  - 1M<n<10M

ImageNet-1K-Camera

Per-image camera parameter annotations for the full ImageNet-1K dataset (1,000 training classes + the 50,000-image validation split, ~1.35M images, 0 failures), predicted by the Puffin camera-centric multimodal model (Qwen2.5-7B + C-RADIOv3-H visual encoder).

Format

The archive mirrors the source ImageNet WebDataset layout: one .tar per source shard (n*.tar per class, plus val.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": 0.0123, "pitch": -0.0871, "vfov": 1.0123, "k1": -0.012, "parse_ok": true}

Camera-parameter distributions

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

Train (1,278,950 images)

Val (49,924 images)

split	roll μ / med / σ	pitch μ / med / σ	FoV μ / med / σ
train	0.1° / 0.0° / 7.8°	−7.8° / −3.7° / 15.7°	28.9° / 25.6° / 8.9°
val	0.1° / 0.0° / 8.5°	−8.8° / −4.5° / 16.7°	30.3° / 27.5° / 9.1°

Reading the distributions

• Roll is sharply peaked at 0° — ImageNet photos are overwhelmingly shot upright/level, with only a small tilted tail.
• Pitch is centered slightly negative (≈ −4° to −8°), i.e. cameras tend to look marginally downward at the subject, with a noticeably wider spread than roll.
• FoV is concentrated in the 20–40° range (median ≈ 26–28°), typical of everyday / mild-telephoto photography; a long thin tail reaches wide-angle.
• Train vs val agree closely on all three parameters, indicating consistent capture statistics across the two splits.

Notes

• Annotations are model predictions, not ground truth.
• Generated with greedy decoding (deterministic) at 640px input.
• Figures reproduce scripts/annotation/stat_camera_captions.py from the Puffin repo (read directly from the packed tars).