Out-of-Distribution Detection using DNN Latent Representations Uncertainty

CEA-LSEA package for Out-of-Distribution (OoD) detection using the uncertainty (entropy) from DNN latent representations. The package has been used with the following applications, the corresponding DNN architectures and datasets:

• Simple Classification:

  • In-Distribution Dataset: CIFAR10

  • Out-of-Distribution Datasets: FMNIST, SVHN, Places365, Textures, iSUN, LSUN-C, LSUN-R

  • DNN Architectures:

    1. ResNet-18

    2. ResNet-18 with Spectral Normalization

• Object Detection:

  • In-Distribution Dataset: BDD100k

  • Out-of-Distribution Datasets: Pascal VOC, Openimages

  • DNN Architectures:

    1. Faster RCNN

• Semantic Segmentation:

  • In-Distribution Dataset: Woodscape & Cityscapes

  • Out-of-Distribution Datasets: Woodscape soiling, Woodscape-anomalies, Cityscapes-anomalies

  • DNN Architectures:

    1. Deeplabv3+

    2. U-Net

In all the above cases, the DNNs were slightly modified to capture epistemic uncertainty using the Monte-Carlo Dropout by adding a DropBlock2D layer.