Cat Royale Comprehensive Dataset

Cat Royale Comprehensive Dataset (CRCD)

The Cat Royale Comprehensive Dataset (CRCD) is a large, multi-modal dataset for feline computer vision research, supporting tasks including instance segmentation, pose estimation, temporal action localisation, and video-text retrieval. The dataset is derived from over 500 hours of 1080p 60fps video footage recorded across 12 days during the Cat Royale artwork and research project, in which a family of three cats lived in a bespoke environment and interacted with a robotic arm. Video was captured from nine cameras positioned throughout the enclosure, yielding a diverse corpus of multi-viewpoint footage featuring a wide range of feline behaviours.

The dataset comprises four subsets:

• CRCD-seg: Over 20,000 images with instance segmentation masks, annotated using Meta’s SAM2 foundation model with human-in-the-loop verification.
• CRCD-pos: Over 5,000 images with 18-point skeletal keypoint annotations for multi-subject cat pose estimation, created using a finetuned DeepLabCut model with human correction and verification.
• CRCD-act: Over 800 video clips annotated with temporally localised action classes drawn from a validated feline ethogram (Nicholson & O’Carroll, 2021), labelled by a qualified clinical animal behaviourist.
• CRCD-cap: Over 800 video clips with detailed natural language captions written by the same animal behaviour expert, describing cat activities, interactions, and behaviours in free-form text.