RGB-D Rat Behavioral Dataset

RGB-D rat behavioral dataset, that is suitable for training automatic behavior recognition in rodents.

The RGB-D frames were recorded using the Microsoft Kinect v2 camera, with 1920x1080 color and 512x424 depth pixel resolution, respectively. They were recorded at a maximum of 30 frames/second (fps) (in low light conditions this value drops to 15fps, typically.
The produced RGB-D dataset was compiled from videos and respective behavior annotations that capture freely-walking Wistar rats in an open-field arena.
The dataset is composed of several ˜10/15 minutes RGB-D video sequences of individual rat behavior. Three different lighting conditions were used to recreate the typical light setups used in behavioral recordings: dim red-light, dim white-light and in total darkness.

The full dataset consists of 24 videos, with a total of 6 hours, divided into two sections: 4 hours of fully annotated sequences (for supervised ML methods; ~180 000 annotated frames) and 2 additional hours of raw behavioral sequences (adding data for unsupervised machine learning methods).
Every RGB-D video frame in the annotated dataset was manually labeled by researchers with experience in ethology, with one of the seven mutually exclusive rat behavioral labels: standstill, moving exploration, local exploration, walking, supported reating, unsupported rearing, grooming.

If you use this software, please acknowledge our work published in: “Gerós, A., Magalhães, A. & Aguiar, P. Improved 3D tracking and automated classification of rodents' behavioral activity using depth-sensing cameras. Behav Res (2020), doi: 10.3758/s13428-020-01381-9”