Consumer-to-Shop Clothes Retrieval Results

[Liu Chang]

Hi, I have problems in understanding the results of clothes retrieval (Table 5 in paper or Table 9 in README). Taking the "class" with ground truth detection box as example, all the results for different validation subsets are much higher than the overall top-20 accuracy (except when zoom_in is large). I understand it is not a simply averaging, but why there is a large difference between them? Are they calculated using the same validation commercial images?

Thank you very much!