Panoptic Segmentation of Individual Pigs for Posture Recognition.

Area of Science:

• Animal behavior science
• Computer vision
• Deep learning

Background:

• Automated recognition systems simplify pig behavioral research.
• Computer vision allows non-intrusive animal evaluation.
• Deep learning methods show promise but have limitations in precise animal contour tracing.

Purpose of the Study:

• To develop a pixel-accurate segmentation method for individual pigs using panoptic segmentation.
• To improve information capture beyond bounding boxes and keypoints for pig behavior analysis.

Main Methods:

• A framework utilizing a neural network for semantic segmentation was developed.
• The framework incorporates specialized network heads and postprocessing techniques.
• A custom dataset of 1000 hand-labeled images was used for testing.

Main Results:

• The panoptic segmentation method achieved a high F1 score of approximately 95%.
• The system demonstrated robustness against occlusions and dirty lenses.
• Pixel-accurate segmentation provides richer data compared to bounding boxes or keypoints.

Conclusions:

• Panoptic segmentation offers a significant advancement for automated pig behavior analysis.
• The developed framework enhances the precision and completeness of animal recognition.
• This technology has the potential to greatly simplify and improve behavioral research in pigs.