📄 Paper: COSTARR (ICCV 2025)
COSTARR uses feature attenuation to generate a COSTARR score which normalizes the scaled prediction logit for Open-Set Recognition.
- Prepare extracted logit and feature vector numpy (.npy) files according to the structure and naming in COSTARR_Driver for each Known/Unknown Dataset you will test. Update the Extractions folder in COSTARR_Driver.py to auto-detect these files.
- Modify the weights in COSTARR_Driver.py to load weights from your target architecture.
- Running COSTARR_Driver.py will produce a system/architecture-named folder in COSTARR_Experiments/ containing output numpy files (.npy) in the following format per row: [gt, predition, score].
COSTARR.py contains the class-level implementation of COSTARR which can be integrated into other projects.