One-Shot Learning with Pseudo-Labeling for Cattle Video Segmentation in Smart Livestock Farming.

Area of Science:

• Computer Vision
• Precision Livestock Farming
• Deep Learning

Background:

• Accurate animal segmentation is crucial for vision-based livestock monitoring and welfare management.
• Traditional deep learning methods require large, pixel-labeled datasets, which are time-consuming to create for animals with irregular shapes and postures.

Purpose of the Study:

• To develop an efficient animal segmentation method for videos that minimizes the need for labeled images.
• To improve automatic animal monitoring and welfare assessment in precision livestock farming.

Main Methods:

• A one-shot learning approach utilizing a pseudo-labeling strategy with an Xception-based Fully Convolutional Neural Network (Xception-FCN) and a pseudo-labeling (PL) module.
• The Xception-FCN model learns features from a single labeled frame, and the PL module refines segmentation using its own predictions.

Main Results:

• Achieved a mean intersection-over-union score of 88.7% and a contour accuracy of 80.8% on a feedlot cattle dataset.
• Outperformed existing state-of-the-art methods like OSVOS and OSMN in cattle video segmentation.

Conclusions:

• The proposed one-shot learning approach effectively segments animals in videos with minimal labeled data.
• This method offers a promising solution for livestock farming applications requiring segmentation and detection.