Input size Aspect Ratio

[ishan]

I am trying to detect objects such as cardboard boxes and people. For most of my training experiments I am using a width of 576 and height of 320 making the aspect ratio: 1.8.

But, I am noticing peoplenet model has an input size dimension of 544 width x 960 height making the aspect ratio 0.56.

I understand that aspect ratio makes sense because people will have a smaller width in comparison to the height; whereas my aspect ratio does the opposite - it has the input width of almost the height. Will this impact my performance for people detection because of the mismatch between the aspect ratio?

[Morganh]

Please note that the peoplenet is trained by 960(width) * 544(height) images instead of 960(height) * 544(width) images.

Your images’ aspect ratio is similar to it.