StreetHazards is a synthetic image dataset for outlier detection, created by researchers using Unreal Engine and the CARLA simulation environment. The researchers inserted a variety of foreign objects into the driving scene and re-rendered the scene with these new objects. There are 5,125 images and semantic segmentation ground truth groups for training, 1,031 groups without outliers for verification, and 1,500 groups with outliers for testing. The training set has 12 categories, namely background, road, street line, traffic sign, sidewalk, pedestrian, vehicle, building, wall, pole, fence, vegetation.

Citation

If you find this useful in your research, please consider citing: @article{hendrycks2019anomalyseg, title={Scaling Out-of-Distribution Detection for Real-World Settings}, author={Hendrycks, Dan and Basart, Steven and Mazeika, Mantas and Zou, Andy and Kwon, Joe and Mostajabi, Mohammadreza and Steinhardt, Jacob and Song, Dawn}, journal={ICML}, year={2022} }