Enabling Effective OOD Detection via Plug-and-Play Network for Mobile Visual Applications

Mobile devices have increasingly integrated with numerous deep learning-based visual applications, such as object classification and recognition models. While these models perform well in controlled environments, their effectiveness declines in real-world environment due to out-of-distribution (OOD) data not seen during training. Existing methods for detecting OOD data often compromise normal data recognition and require extensive training on unattainable OOD data. To address these issues, we propose $\mathtt {POD}$, a framework designed to enhance mobile visual applications by providing high-precision OOD detection without affecting original model performance. In the offline phase, $\mathtt {POD}$ generates OOD detectors from any classification model by analyzing model’s neuron responses to various data types. In the online phase, it continuously adjusts decision boundaries by integrating results from both the original model and the detector. Evaluated on two public datasets and one self-collected dataset across various popular classification models, $\mathtt {POD}$ significantly improves OOD detection performance while maintaining the accuracy of original models.