This paper proposes ODIN, a simple and effective method for detecting out-of-distribution (OOD) images in neural networks. ODIN does not require retraining the pre-trained network and is compatible with various architectures and datasets. The method leverages temperature scaling and input perturbations to enhance the separation of softmax score distributions between in-distribution and OOD images, leading to improved detection performance. Experiments show that ODIN significantly outperforms the baseline method (Hendrycks & Gimpel, 2017), reducing the false positive rate (FPR) from 34.7% to 4.3% on the DenseNet model when the true positive rate (TPR) is 95%. ODIN is tested on multiple datasets, including TinyImageNet, LSUN, Gaussian noise, and uniform noise, demonstrating its effectiveness across diverse scenarios. The method is also shown to be robust to parameter settings and generalizable across different validation datasets. Theoretical analysis supports the effectiveness of temperature scaling and input preprocessing in improving detection performance. The paper concludes that ODIN provides a significant improvement in OOD detection without requiring changes to the pre-trained network.This paper proposes ODIN, a simple and effective method for detecting out-of-distribution (OOD) images in neural networks. ODIN does not require retraining the pre-trained network and is compatible with various architectures and datasets. The method leverages temperature scaling and input perturbations to enhance the separation of softmax score distributions between in-distribution and OOD images, leading to improved detection performance. Experiments show that ODIN significantly outperforms the baseline method (Hendrycks & Gimpel, 2017), reducing the false positive rate (FPR) from 34.7% to 4.3% on the DenseNet model when the true positive rate (TPR) is 95%. ODIN is tested on multiple datasets, including TinyImageNet, LSUN, Gaussian noise, and uniform noise, demonstrating its effectiveness across diverse scenarios. The method is also shown to be robust to parameter settings and generalizable across different validation datasets. Theoretical analysis supports the effectiveness of temperature scaling and input preprocessing in improving detection performance. The paper concludes that ODIN provides a significant improvement in OOD detection without requiring changes to the pre-trained network.