A Appendix

Figure 5 shows the visualization of coreset in the memory bank. Though replacing the greedy sampling with random sampling, PatchCore avoids most noisy features but is poor at model training set and still misled by some noise. The coreset of SoftPatch is clean and decentralized. We use t-SNE for dimension reduction for visualization. SoftPatch wipe off the noisy patch and model the nominal data properly.