MCB-Oriented RCNN: A Lightweight Rotated Detector for 6-DoF Grasping in Solid Waste Sorting

Abstract

Visual sorting robots often struggle with conventional horizontal bounding boxes when detecting irregularly placed solid waste. These boxes exhibit excessive background noise, making it difficult to extract precise depth data for 3D pose estimation. To address this issue, we developed MCB-Oriented RCNN, a more efficient detector for rotated objects based on the Oriented-RCNN framework. We replaced the resource-intensive ResNet-50 backbone with MobileNet-V3 to increase inference speed and integrated BiFPN and Coordinate Attention to maintain localization accuracy. Since lightweight models often suffer from accuracy losses, a teacher-student distillation strategy was applied to restore performance. Our experiments on a custom waste dataset show that the model achieves an mAP of 97.16% at 43.0 FPS. This corresponds to a 74.8% speed increase over the baseline model with a negligible accuracy loss of only 0.74%. These rotated boxes provide a direct approach to computing the 6-DoF coordinates required for robotic grasping.