RGB-D Image Dataset for Automated Body Condition Scoring of Dairy Cows in Rotary Milking Parlors

1. Overview and Industrial Context

The RGB-D Image Dataset for Automated Body Condition Scoring is an industrial-scale repository designed to facilitate the development and benchmarking of computer vision models for precision livestock farming. Collected at a commercial dairy facility in the Saraktashsky district, Orenburg region, Russian Federation (Coordinates: 51.928054, 56.006392), this dataset addresses the specific challenges of automated monitoring within high-throughput rotary milking systems. It provides a robust baseline for real-time BCS estimation under varying lighting conditions and animal postures typical of industrial carousel operations.

2. Population Metadata and Animal Management

The dataset represents a diverse biological cohort, ensuring high external validity for commercial dairy operations:

• Sample Size: The repository comprises 25,700 synchronized RGB-D image pairs derived from 1,025 unique animals.

• Breed Composition: To ensure morphological diversity, the population includes Holstein, Red Steppe, Jersey, and various crossbreeds.

• Lactation Status: All subjects were in the active lactation phase at the time of acquisition, spanning multiple age groups and lactation orders.

• Management Regime: Animals were maintained under a standardized feeding and management protocol, with data captured passively during the routine milking cycle to minimize handling-induced stress.

3. Hardware Configuration and Acquisition Protocol

The sensing pipeline utilized a Microsoft Kinect RGB-D sensor to capture synchronized textural and geometric data:

• Sensor Geometry: The camera was mounted in a fixed nadir (overhead) position at a height of approximately 3.05 meters. This height was optimized to encompass the full width of the transition zone and the entire pelvic-lumbar region of the bovine subjects.

• Temporal Resolution: Images were captured at a constant frame rate of 30 FPS. Acquisition was performed three times daily (morning, afternoon, and evening sessions) to capture physiological snapshots across different stages of the daily production cycle.

4. Expert Annotation and Ground-Truth Methodology

Reference labels were established through manual assessment by trained veterinary specialists following the Ferguson 5-point protocol. Labels are provided in 0.25-point increments, ranging from 2.75 to 5.00.

5. Dataset Statistics and Class Distribution

The dataset exhibits a distribution characteristic of well-managed commercial herds, with a primary concentration in the optimal health range:

• Target Concentration: The majority of samples fall within the 3.25 (37.04%) and 3.50 (28.09%) range.

• Extreme Condition Classes: While mid-range scores predominate, the dataset includes sufficient representations of leaner (BCS 2.75–3.00) and over-conditioned (BCS 4.25–5.00) animals to support robust tail-distribution learning.

6. File Structure and Data Organization

The repository is organized according to the following directory structure:

• /data/frame/rgb/: This directory contains visual spectrum images stored in standard .png format.

• /data/frame/depth/: This directory comprises geometric depth maps, normalized to an 8-bit grayscale .png format.

• Dataset.xlsx: A comprehensive metadata file providing unique animal identifiers, frame, data, and expert-assigned Body Condition Scores (BCS), RGB-path and Depth-path.