Improvement of Detection Accuracy by Additional Judgments using LightGBM on YOLOX Model Detection Results

Abstract

In object detection tasks, users generally can adjust the confidence score threshold to control the balance between false positives and missed detections. Bounding boxes with confidence scores exceeding this threshold are included in the final detection results. However, this creates a trade-off: raising the threshold reduces false positives but increases missed detections, while lowering the threshold has the opposite effect. This trade-off poses a challenge, as threshold adjustment alone is insufficient to effectively address both issues simultaneously. To overcome this limitation, this study proposes a hybrid model called GBDT-YOLOX, which combines a Gradient Boosting Decision Tree (GBDT) model with the detection results produced by YOLOX, one of the state-of-the-art object detection model. The GBDT model refines the raw detection outputs by leveraging additional contextual and feature-level information, enabling more informed decisions and reducing reliance on a single threshold. Experimental results on the COCO dataset demonstrate that GBDTYOLOX achieves 1.5% and 3.0% improvements in Average Precision (AP) and Average Recall (AR), respectively. These results indicate that the proposed method effectively reduce both false positives and missed detections. Additionally, to evaluate its performance in a more challenging domain-specific scenario, experiments were conducted on the RailSem19 dataset, which focuses on railway environments with unique detection challenges. The results showed an approximate 3% improvement in AP, further validating the effectiveness of GBDT-YOLOX in reducing detection errors in complex and specialized applications.