Abstract
While improving the performance on the out-of-distribution (OOD) benchmark dataset, the existing approach misses a portion of the valid discriminative information such that it reduces the performance on the same manifold OOD (SMOOD) data. The key to addressing this problem is to prompt the model to learn effective and comprehensive in-distribution (ID) semantic features. In this paper, two strategies are proposed to improve the generalization ability of the model to OOD data. Firstly, the original samples are replaced by features extracted from multiple "semantic perspectives" to obtain a comprehensive semantics of the samples; Second, the mean and variance of the batch samples are perturbed in the inference stage to improve the sensitivity of the model to the OOD data. The method we propose does not employ OOD samples, uses no pre-trained models in training, and does not require pre-processing of samples during inference. Experimental results show that our method enhances the semantic representation of ID data, surpasses state-of-the-art detection results on the OOD benchmark dataset, and significantly improves the performance of the model in detecting the SMOOD data.