Uncertainty-Aware and objectness-Guided Feature Quantization for Robust Object Detection

Object detection models face challenges in accurately extracting features due to complex backgrounds, small objects, and environmental noise. Particularly, noise present in images increases pixel-level uncertainty, blurring the boundaries between objects and backgrounds, thereby causing confusion in object information. This increased uncertainty in specific regions reduces the reliability of feature representations, significantly deteriorating detection performance. To address these issues, this paper proposes a novel feature quantization framework that simultaneously utilizes uncertainty and objectness information. The proposed method independently estimates uncertainty maps and objectness maps at the pixel level to effectively identify uncertain regions, subsequently training a vector quantization codebook based on features from high-objectness regions. During inference, regions with both high uncertainty and high objectness are treated as objects, with their features replaced by the nearest code vector to restore object information accurately. Conversely, regions with high uncertainty but low objectness are considered background noise and are masked by setting their features to zero, thus minimizing the probability of errors. Experimental evaluations using a YOLO-based object detection model on the BDD100K and KITTI dataset demonstrate improvement in mAP performance compared to the conventional YOLOv12 model.