CondorG/LambingSheep

tags:

• computer-vision
• pose-estimation
• animal-behavior
• sheep
• keypoint-detection
• coco-format

Dataset Description

LambingSheep is a specialized pose estimation dataset for periparturient (around lambing) sheep, constructed to support automated monitoring of prepartum behaviors in livestock breeding environments. The dataset captures 17 anatomical keypoints of ewes during critical prepartum stages, enabling fine-grained behavior analysis such as standing, lying, walking, pawing, standing up, and lying down transitions.

Dataset Summary

Split	Images
Train	1,417
Validation	175
Test	183
Total	1,775

Collection environment

• Location: Lambing shed of Gansu Qinghuan Mutton Sheep Breeding Co., Ltd.
• Period: June 14–20, 2024
• Pen Specifications: Approx. 7 m × 5 m per pen, housing 10–15 pregnant ewes
• Flooring: Grid-patterned plastic slatted flooring for manure dropping and hygiene maintenance

Collecting device

Parameter	Configuration
Camera	TP-LINK TL-IPC48AW network camera
Storage	256 GB SD card (local storage)
Total Raw Data	140 hours / 215 GB
Mounting	Brackets at 0.1 m horizontal distance from pen fence
Height	1.1 m above ground
Coverage	Full panoramic pen coverage after pitch angle adjustment

Sampling Strategy

Keyframes were obtained by trimming 2-second segments of six prepartum behaviors and applying adaptive sampling via FFmpeg.

Behavior Category	Behaviors	Sampling Rate (frames/2s segment)	Rationale
Static postures	Standing, Lying	1 frame	Minimal motion, low dynamic complexity
Periodic motions	Walking, Pawing	3 frames	Regular cyclic movement
Significant deformations	Standing up, Lying down transition	6 frames	Large shape changes require dense sampling