Figure 5
From: Improve the performance of CT-based pneumonia classification via source data reweighting
Randomly-sampled source images whose importance weights learned by different methods are close to 0. Our method can successfully identify images containing artifact noises such as bounding boxes or having large domain discrepancies with target data in terms of appearance, texture, color, scale, etc. In contrast, OML and MTL incorrectly assign close-to-zero weights to some images that are clean and have large domain similarity to target data.
