Paper Title
Automatic Segmentation of Birds Using a Combination of Object Detection and Foundation Image Segmentation Models

Abstract
This study introduced an innovative method for automatic bird segmentation by combining an object detection model (i.e., YOLOv7) with a foundation image segmentation model (i.e., Segment Anything Model, SAM). YOLOv7 detected individual birds in images and calculated bounding box prompts of each detected bird for the SAM, enabling detailed and efficient segmentation without manual point inputs. The developed method was compared with various segmentation methods, including YOLOv8, Thermal image + MobileSAM, Thermal image + SAM, Thermal image + FastSAM, Mask R-CNN, and YOLOv7 (providing centroids of detected birds as point prompts) + SAM. The results showed that the proposed method outperformed all of the comparative segmentation methods, with the highest precision of 92.5%, recall of 98.2%, F1 score of 95.1%, IoU of 91.0%, and success rate of 98.0%. The study highlights a significant advancement in automatic image segmentation techniques with less intensive human annotation than standard deep learning-based image segmentation methods. The developed methods can be scaled up and transferred to various agricultural, environmental, medical, geographical, and urban planning applications. Keywords - Poultry, Segment Anything Model, YOLOv7, Segmentation