<![CDATA[Robotic bin-picking requires accurate detection of randomly oriented objects under cluttered conditions. Conventional axis-aligned bounding boxes struggle to distinguish adjacent objects, motivating the use of oriented bounding boxes (OBB). This paper proposes a complete pipeline for bin-picking object detection that combines the Segment Anything Model 2 (SAM2) with YOLOv11-OBB. A three-stage auto-annotation pipeline first applies a YOLOv11s horizontal bounding-box detector to localize each object and assign its class label. SAM2 then performs automatic instance segmentation within each detected bounding-box region without requiring manual point prompts. Last, the resulting masks are converted to OBB annotations via minimum-area rectangle fitting, reducing annotation time by approximately 877× compared with manual labeling. YOLOv11-OBB featuring C2PSA attention, C3k2 convolution blocks, and an anchor-free rotated detection head is trained for 300 epochs on a purpose-built dataset of three cylindrical object classes (white, black, and red) captured in a UR3 collaborative-robot workspace. Experiments demonstrate an overall mAP@0.5 of 0.995 and mAP@0.5:0.95 of 0.949, with an inference time of 138 ms per frame on a consumer CPU. The results indicate that the proposed pipeline is well-suited for oriented object detection in industrial bin-picking applications.]]>
Copyrights © 2026
