Label Refinement for Robot Vision Object Detection via Dynamic Threshold Estimation and Optimal Refinement Model Selection

Abstract

This study proposes a label refinement method for robot vision object detection utilizing multi-model-based dynamic threshold estimation. The ultimate goal of this research is to improve the recognition accuracy and robustness of intelligent waste-sorting robotic arm vision systems. To overcome the limitations of conventional fixed-threshold approaches, the proposed method applies multiple dynamic adaptive confidence thresholds that adjust according to the data distribution. Furthermore, it enhances refinement accuracy through an adaptive geometric prior-based noise reduction technique. Specifically, the proposed method adopts a two-stage approach, starting with an initial estimation of the label data to be refined. This is followed by the selection of an optimal refinement model, which collectively enhances both the precision and processing speed of the refinement process. Experimental results demonstrate that the proposed method improves the inference speed while increasing mean Average Precision (mAP) @.5:.95 by 0.0282 compared to the conventional methods. These findings indicate that the proposed framework effectively corrects inherent labeling errors and suppresses catastrophic forgetting during the fine-tuning process. Ultimately, this advancement provides a highly reliable and generalizable vision model solution for practical deployments in real-world automated sorting environments.