VOS: Learning What You Don't Know by Virtual Outlier Synthesis

Authors: Xuefeng Du, Zhaoning Wang, Mu Cai, Yixuan Li

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental VOS achieves competitive performance on both object detection and image classification models, reducing the FPR95 by up to 9.36% compared to the previous best method on object detectors. Code is available at https://github.com/deeplearning-wisc/vos.
Researcher Affiliation Academia Xuefeng Du, Zhaoning Wang, Mu Cai, Yixuan Li Department of Computer Sciences University of Wisconsin Madison {xfdu,mucai,sharonli}@cs.wisc.edu
Pseudocode Yes Algorithm 1 VOS: Virtual Outlier Synthesis for OOD detection
Open Source Code Yes Code is available at https://github.com/deeplearning-wisc/vos.
Open Datasets Yes We use PASCAL VOC (Everingham et al., 2010) and Berkeley Deep Drive (BDD-100k) (Yu et al., 2020) datasets as the ID training data. For both tasks, we evaluate on two OOD datasets that contain subset of images from: MS-COCO (Lin et al., 2014) and Open Images (validation set) (Kuznetsova et al., 2020).
Dataset Splits Yes We use CIFAR-10 (Krizhevsky & Hinton, 2009) as the ID training data, with standard train/val splits. ID val dataset VOC val BDD val
Hardware Specification Yes We run all experiments with Python 3.8.5 and Py Torch 1.7.0, using NVIDIA Ge Force RTX 2080Ti GPUs.
Software Dependencies Yes We run all experiments with Python 3.8.5 and Py Torch 1.7.0, using NVIDIA Ge Force RTX 2080Ti GPUs.
Experiment Setup Yes We use the Detectron2 library (Girshick et al., 2018) and train on two backbone architectures: Res Net-50 (He et al., 2016) and Reg Net X-4.0GF (Radosavovic et al., 2020). We employ a two-layer MLP with a Re LU nonlinearity for φ in Equation 5, with hidden layer dimension of 512. For each in-distribution class, we use 1,000 samples to estimate the class-conditional Gaussians. [...] The PASCAL model is trained for a total of 18,000 iterations, and the BDD-100k model is trained for 90,000 iterations. We add the uncertainty regularizer (Equation 5) starting from 2/3 of the training. The weight β is set to 0.1.