CattleFace-RGBT: Cattle Facial Landmark Dataset with RGB-Thermal Imagery

Dataset Description

CattleFace-RGBT is the first publicly available multimodal dataset featuring paired frontal-view RGB and thermal facial images of cattle, annotated with 13 facial keypoints and associated ground-truth rectal temperature measurements. The dataset was developed to support research in automated cattle fever estimation and precision livestock farming.

Paper: CattleFever: An automated cattle fever estimation system Published in: Smart Agricultural Technology, Volume 12, 2025

Dataset Summary

Component	Count
RGB images (annotated)	1,890
Thermal JPG images (annotated, colorized)	2,611
Raw thermal TIFF frames	30,954
Thermal videos (.mp4)	51
Unique cattle	108
Cattle with temperature readings	21
Facial keypoints per image	13
Recording dates	3 (Feb 1, Feb 6, Feb 13)

Dataset Structure

CattleFace-RGBT/
├── README.md
├── rgb/                          # RGB images organized by folder
│   ├── 1/
│   ├── 17/
│   ├── 25/
│   ├── 50/
│   └── 64/
├── thermal/                      # Colorized thermal JPG images
│   ├── 1/
│   ├── 2/
│   ├── 17/
│   ├── 25/
│   ├── 50/
│   └── 64/
├── thermal_raw/                  # Raw thermal TIFF frames (temperature data)
│   ├── 02_01/                    # Feb 1 recording session
│   ├── 02_06/                    # Feb 6 recording session
│   └── 02_13/                    # Feb 13 recording session
└── annotations/
    ├── rgb_keypoints.json        # COCO-format keypoint annotations for RGB
    ├── thermal_keypoints.json    # COCO-format keypoint annotations for thermal
    ├── metadata.csv              # Cow ID, temperature, and data mapping
    └── cow_mapping.json          # Sequence number → cow tag ID mapping

Annotation Format

Annotations follow the COCO keypoint format:

Images

{
  "id": 0,
  "file_name": "rgb/1/00001.jpg",
  "width": 2560,
  "height": 1440,
  "folder": "1",
  "frame_id": "00001"
}

Annotations

{
  "id": 0,
  "image_id": 0,
  "category_id": 1,
  "keypoints": [x1, y1, v1, x2, y2, v2, ...],
  "num_keypoints": 13,
  "bbox": [x, y, width, height],
  "area": 123456,
  "iscrowd": 0
}

13 Facial Keypoints

Index	Name	Description
1	left_ear_base	Base of left ear
2	left_ear_middle	Middle of left ear
3	left_ear_tip	Tip of left ear
4	poll	Top of head (poll)
5	right_ear_base	Base of right ear
6	right_ear_middle	Middle of right ear
7	right_ear_tip	Tip of right ear
8	left_eye	Left eye
9	right_eye	Right eye
10	muzzle	Center of muzzle
11	left_nostril	Left nostril
12	right_nostril	Right nostril
13	mouth	Mouth

Visibility flag: 0 = not labeled, 2 = labeled and visible.

Raw Thermal Data

The thermal_raw/ directory contains raw TIFF frames from the ICI FMX 400 thermal camera (384 x 288 pixels). Each pixel contains a temperature value in Celsius. These files can be read with:

from PIL import Image
import numpy as np

tiff = Image.open("thermal_raw/02_01/0001_Video_Frame_1.tiff")
temp_array = np.array(tiff)  # Temperature values in Celsius

TIFF filenames follow the pattern: {sequence_num}_Video_Frame_{frame_num}.tiff

Use cow_mapping.json to map sequence numbers to cow tag IDs and temperatures.

Temperature Data

Ground-truth rectal temperatures (in Fahrenheit) are available for 21 cattle across 3 recording sessions. The mapping is provided in metadata.csv and cow_mapping.json.

Data Collection

Data was collected at the Arkansas Agricultural Experiment Station, Savoy Research Complex, Beef Cattle Research Area, in partnership with the University of Arkansas. The setup used:

• RGB camera: Standard webcam (2560 x 1440 resolution)
• Thermal camera: ICI FMX 400 (384 x 288 pixel resolution, 50 Hz frame rate, < 0.03°C thermal sensitivity)
• Temperature: Rectal thermometer (ground truth)

Each calf was guided into a cattle squeeze chute for ~20 seconds of synchronized RGB and thermal video recording.

Supported Tasks

1. Cattle facial landmark detection — Detect 13 keypoints on cattle faces
2. Cattle face detection — Detect and localize cattle faces using bounding boxes
3. Core body temperature estimation — Predict rectal temperature from thermal facial features

Recommended Splits

As described in the paper:

• Keypoint detection: 70% train / 30% test (random split)
• Temperature estimation: 80% train / 20% test

Citation

@article{pham2025cattlefever,
  title={CattleFever: An automated cattle fever estimation system},
  author={Pham, Trong Thang and Coffman, Ethan and Kegley, Beth and Powell, Jeremy G. and Zhao, Jiangchao and Le, Ngan},
  journal={Smart Agricultural Technology},
  volume={12},
  pages={101434},
  year={2025},
  publisher={Elsevier},
  doi={10.1016/j.atech.2025.101434}
}

License

This dataset is released under the CC BY 4.0 license.

Contact

For questions about this dataset, please contact:

• Trong Thang Pham (tp030@uark.edu) — AICV Lab, University of Arkansas