problem with test config

[khanhnn00]

Hello authors,
thanks for your dedicated work. However, I am currently confused with the FLIP and SCALES options in testing phase. I do not understand how it affects my final result, because I'm running on a single test image only. Do I have to keep the default config? I'm setting it as False in FLIP and only 1 in SCALES only, and the results on random images indicates it does not much differ from your default config.

Thanks in advance.

[Nikita Araslanov]

Hi @khanhnn00,
these settings are for multi-scale inference, which is standard in this domain. SCALES specifies the scaling factor, while FLIP indicates if the images should be flipped. For example if SCALES: [1, 0.5, 1.5, 2.0] and FLIP: True, at inference time the network gets a batch with 8 versions of the same image (four scales + four scales flipped). The final prediction is a merge of the predictions from this batch.
Please, do look in the code if you change the number of scales or disable the flipping, as these settings are not "plug-and-play" in the current implementation, which expects this multi-scale setup.

Best,
Nikita