CenterNet image_resizer Strange Performance

[etatbak]

I train my custom model with CenterNet HourGlass104 1024x1024 My dataset contains only images with 720:1280 size. In config file when I change keep_aspect_ratio_resizer as 720:1280 it doesn't detect accurately. But when I set aspect_ratio as 1280:1280 or 1024:1024 it works much better.
Someone please can explain why I get better results with these scales even I have lower size images? Here a part of my config file:

image_resizer {
  keep_aspect_ratio_resizer {
    min_dimension: 720
    max_dimension: 1280
    pad_to_max_dimension: true
  }
}

[etatbak]

Hi @pkulzc maybe you have an idea about this issue?