CatVision: Human–Cat Vision Frame Pairs

Official dataset for the paper:

Purrturbed but Stable: Human-Cat Invariant Representations Across CNNs, ViTs and Self-Supervised ViTs
Arya Shah et al. · arXiv:2511.02404

This dataset contains 346,400 paired video frames rendered under human vision and simulated cat vision optics. It was used to benchmark cross-species representational alignment across CNNs, supervised ViTs, windowed transformers, and self-supervised ViTs (DINO) using CKA and RSA.

Dataset Structure

Column	Type	Description
pair_id	string	Unique identifier for each frame pair
video_id	string	Source video (video1–video191)
frame_filename	string	Original frame filename
human_frame	image	Frame as seen by humans
cat_frame	image	Frame simulated for Felis catus vision

Statistics

Stat	Value
Videos	191
Frame pairs	346,400
Format	Parquet (1 file / video, Snappy)
Image encoding	JPEG

Usage

from datasets import load_dataset

ds = load_dataset("aryashah00/CatVision")
sample = ds["train"][0]

# Access images (returned as PIL.Image)
human_img = sample["human_frame"]
cat_img    = sample["cat_frame"]

Citation

@misc{shah2024purrturbed,
  title     = {Purrturbed but Stable: Human-Cat Invariant Representations
                Across CNNs, ViTs and Self-Supervised ViTs},
  author    = {Arya Shah and others},
  year      = {2024},
  eprint    = {2511.02404},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url       = {https://arxiv.org/abs/2511.02404}
}

License

Released under CC BY 4.0.