multi-scale testing?

[zimenglan-sysu-512]

recently i want to try multi-scale testing. since the rpn network does not support multi-scale or batchsize > 1, so i set batchsize = 1 and loop for multi-scale to stack the results of each scale for one image. but i got some wired problems, when i run the demo.py: if try different scale (only one scale at the same time), the results look normal, for example:

scale: 684

scale: 564

scales: 564 and 684 together

it is so wired that the results from mulit-scales have some strange bounding boxes. but if try single scale multiply times for the same image, results look normal; if i try multi-scale case, i found that the the results of last scale in cfg.TEST.SCALES look normal, the rest from other scales but not the last one look wired. if try to replicate the network net, that is to say each scale has its own net, the results stacked from these scales look normal.

i try to figure out what happens, but can't find any useful info, i guess the network net in the multi-scale testing get dirty.

have anyone met this before?

thanks.

[dongzhuoyao]

I recently have a research in image segmentation,it also says "multi-scale testing",so what is "multi-scale testing" mean?

[feymanpriv]

Have you solved? I met the same problem.

[hzh8311]

@zimenglan-sysu-512 Did you trained the networks on a multi-scale way or you just trained the network on single scale?

[wwwpbai]

@hzh8311 , Do you know how to achieve multi-sacle training now?