Synthetic Harmless outlier Images generator From Training samples (SHIFT)

An offitial implementation for "Improving Out-of-Distribution Detection Performance using Synthetic Outlier Exposure Generated by Visual Foundation Models (BMVC 2023)".

Authors: Gitaek Kwon, Jaeyoung Kim, Hongjun Choi Byungmoo Yoon, Sungchul Choi, Kyuhwan Jung

Runs

STEP 1: Generate OOD Samples using SHIFT

# generate OOD samples from STL10 dataset.
python generate_ood.py --dataset stl10

STEP 2: Train the rejection network

# STL10 (ID) vs SVHN (OOD)
python train.py --dataset stl10

Citation