Training-Free OOD Object Detection Leveraging Pre-Trained Segmentation Model Competency

This paper addresses a method for detecting Out-of-Distribution (OOD) objects by leveraging pre-trained segmentation models. The research on OOD detection focuses on the ability to accurately identify and classify untrained classes as “unknown.” Previous methods rely on strong augmentation to detect OOD objects. However, these augmentation-based methods assume specific characteristics of OOD objects and can suffer from overfitting due to the limited datasets. In this study, we propose a training-free OOD object detection approach that infers OOD objects from a pre-trained segmentation model without additional training. Specifically, our method combines two modules: rejection of regions confidently recognized by the model (inliers) and selection of masks that capture the characteristics of OOD objects (outliers). Furthermore, using a class-agnostic segmentation model, the probability of OOD objects is refined at the object level, enhancing performance. Our method shows competitive performance on datasets where OOD objects are encountered in autonomous driving contexts. Additionally, we show the utility of measuring the model's competency to recognize what it knows and doesn't know from the perspective of the pre-trained model.