A Community Detection Approach to Cleaning Extremely Large Face Database

Figure 1

After feature extraction, we delete mislabeled images by detecting communities in the “face similarity graph,” reserving large communities and removing the rest. Finally, we relabel the previously deleted images.