Calibration of the Open-Vocabulary Model YOLO-World by Using Temperature Scaling

Abstract

In many areas of the real world, such as robotics and autonomous driving, deep learning models are an indispensable tool for detecting objects in the environment. In recent years, supervised models such as YOLO or Faster R-CNN have been increasingly used for this purpose. One disadvantage of these models is that they can only detect objects within a closed vocabulary. To overcome this limitation, research is currently being conducted into models that can also detect objects outside the known classes of the training data set. A model is therefore trained with base classes and can recognize novel, unseen classes – this is referred to as openvocabulary detection (OVD). Novel models such as YOLO-World offer a solution to this problem, but they tend to over- or underestimate when calculating confidence values and are therefore often poorly calibrated. However, reliable determination of confidence values is a crucial factor for the use of these models in the real world to ensure safety and trustworthiness. To address this problem, this paper investigates the influence of the calibration method temperature scaling on the OVD model YOLO-World. The optimal T-value is determined by 2 calibration data sets (Pascal VOC and Open Images V7) and then evaluated on the LVIS minival dataset. The results show that the use of temperature scaling improved the Expected Calibration Error (ECE) from 6.78% to 2.31%, but the model still tends to overestimate the confidence values in some bins.