1. download the randomly cropped images from https://studentutsedu-my.sharepoint.com/:f:/g/personal/bo_peng-7_student_uts_edu_au/IgBg6851laF-R4a7v3fVtR3XAU-A3lofQHSgG2PICWFUnfc?e=cFC8bn

2. extract CLIPB/16 features for both original and the cropped images and stroed in ./CLIP-features. For your convenience, the extracted features can be from https://studentutsedu-my.sharepoint.com/:f:/g/personal/bo_peng-7_student_uts_edu_au/IgAXx2_Ed7NiRZ08ZzdbhRhgARm6_baaomqR628P1NLAH90?e=dB7mcz

3. with the precomputed textual features of both ID and negative labels, run main.online.py to reproducce our main results in Table 1


If any of the above links are not accessible, please be free to contact at bo.peng-7@student.uts.edu.au