Submitted to: BOP Challenge 2025

Training data: provided PBR + custom synthetic

Notes: Our cascaded YOLO–YOLACT framework integrates additional networks to maximize recall and improve precision. A Box Reorganization algorithm redefines precise RoIs, followed by an Image Classifier and Feature-Level Refinement to enhance class accuracy. Scene-Level Fusion combines class, score, and box information, yielding robust segmentation of small and heavily occluded objects.