Out-of-Distribution Object Detection in Street Scenes via Synthetic Outlier Exposure and Transfer Learning
SynOE-OD is a method that allows a single object detector to simultaneously find both normal and rare objects in street scenes. It generates synthetic, unusual objects using Stable Diffusion and uses an open-vocabulary detector to create training data. This data is then used for transfer learning to fine-tune the detector, making it robust against out-of-distribution objects without needing separate auxiliary branches during testing.