refineDLC: an advanced post-processing pipeline for DeepLabCut outputs

Description of the data and file structure

Two distinct experimental datasets were compiled to reflect diverse conditions commonly encountered in behavioral and locomotion research:

Walking dairy cattle:

17 high-resolution (1080p) videos recorded under controlled experimental conditions.

Environment: Consistent, controlled, and structured setup ensuring minimal occlusion or visual interference.

Purpose: Validation of pipeline performance under ideal recording conditions.

Trotting horses:

524 publicly available lower-resolution (720p) videos sourced from YouTube (Arabian Essence channel, accessed September–October 2023).

Environment: Uncontrolled conditions with frequent challenges including variable camera angles, occlusions, inconsistent lighting, and irregular camera movements.

Purpose: Robustness testing of the refineDLC pipeline under realistic, highly variable field conditions.

Files and variables

File: data.zip

Description: 

The dataset (/cattle and /horses) includes raw DLC outputs structured as follows:

Row 1: DLC model identifier (scorer name).

Row 2: Annotated body parts (each repeated for x, y, and likelihood values).

Row 3: Coordinate labels indicating horizontal (x), vertical (y), and likelihood (confidence) scores.

Rows 4+: Frame-by-frame numerical data capturing body part positions and tracking likelihood.

Contents

Raw positional coordinates (x, y) for 22 tracked body parts.

Likelihood scores indicating confidence in tracking accuracy.

Video frames processed per species:

Dairy cattle (walking): 17 videos

Horses (trotting): 524 videos

Note: Body parts tracked include anatomical landmarks relevant for locomotion analysis (e.g., withers, croup, elbows, hooves, fetlocks) as well as additional labels included to enhance model training (e.g., handler’s limbs in horse videos).