why is testing image-size different from training

[RuibinMa]

Thanks for sharing the code first! I'm wondering why in the test script you used 1024 as image size and multiscale but in train script you used 362 and no multiscale. Could you give some insights on that?

In general, do you recommend always using the 1024 + multiscale setting in practice?

[filipradenovic]

In general, for inference, 1024 performs the best, and multi-scale helps. However, for training, the benefits of larger scale diminish. You can run training with 1024, and get some small benefit, but at a very high cost of GPU memory and training time.