About lvis dataset - Githubissues

tztztztztz / eqlv2

The official implementation of Equalization Loss v1 & v2 (CVPR 2020, 2021) based on MMDetection. https://arxiv.org/abs/2012.08548 https://arxiv.org/abs/2003.05176

Apache License 2.0

155 stars 22 forks source link

About lvis dataset #5

Closed U-Help closed 3 years ago

U-Help commented 3 years ago

Hello, I'm sorry to bother you but I get into trouble these days. As far as I know, AP is the average of per-category precision. That's same for APr, APc, APf. Therefore, we should have the following equation: However, the statistics are against it. Do you know the reason?

tztztztztz commented 3 years ago

Hi, The overall AP is computed by

$AP_{overall}=\frac{\sum_{j}^{C} AP_{j}}{C}$

we do not group all categories into three groups (r, c, f) then calculate the average. Instead, the AP is directly aggregated on all categories.

Actually, the computation is more complicated, you can refer to the implementation details