Semi-Automatic Labeling with SAM3 and Instance Segmentation of Broiler Chickens Using YOLO26n-seg

Abstract

Instance segmentation is a relevant tool for the automated monitoring of broiler chickens in high-density overhead scenes; however, the manual generation of masks requires considerable effort. This study aimed to develop and evaluate a semi-automatic labeling workflow based on SAM3, automated cleaning, human review, and YOLO26n-seg training. A total of 1.863 images were extracted from 14 videos and distributed by video across the training, validation, and test subsets. SAM3 generated 210.762 polygon annotations, of which 134 duplicates were removed during automated cleaning. The annotations were subsequently reviewed manually and used to train YOLO26n-seg with an input resolution of 960 × 960 pixels. On the validation set, the model achieved a mask precision of 0,78845, a recall of 0,90522, a mAP@0,50 of 0,89476, and a mAP@0,50:0,95 of 0,68705; on the test set, it obtained 0,78639, 0,87559, 0,88387, and 0,67593, respectively. The results showed stable performance across both subsets and a strong ability to recover instances. Nevertheless, accurate contour delineation remained more challenging in regions characterized by high density, physical contact, and partial occlusion. The proposed workflow represents a reproducible alternative for supporting the construction of instance-segmentation datasets in real poultry-farming environments.