Work

Boundary-Aware Semantic Segmentation of Pig Point Clouds in Commercial Housing Environments

Year: 2026
Type: article
Abstract: In commercial pig-housing environments, contact between pig-body point clouds and background structures can produce blurred boundaries, local adhesion, and background misclassification, which may propagate to downstream three-dimensional phenotyping. This study proposes a boundary-aware pig point-cloud segmentation method based on an Octree Transformer backbone. The framework combines multi-scale feature fusion, soft-distance boundary pseudo-labels, and bidirectional semantic-boundary interaction. On the prespecified independent test split, the method achieved an accuracy (Acc) of 98.65%, a mean intersection over union (mIoU) of 97.36%, and a boundary intersection over union (B-IoU of 88.77%. These are single-run point estimates; inferential significance and variability across random initializations were not assessed. The evaluation was intended to examine the complete boundary-aware workflow under a fixed independent animal-level protocol rather than to establish statistical significance from small numerical margins. The observed advantage was largest for B-IoU, consistent with the method’s explicit boundary objective, and should therefore be interpreted together with Acc, mIoU, and qualitative results. The findings support the method under the two sampled farms and fixed acquisition configurations, while cross-farm and cross-device generalization remain to be established.
Language: English
Subfield: Plant Science
Open Access status: hybrid
OpenAlex ID: W7202352361
FWCI:
Cites: 32
Cited by:
Related to: 10