Regarding bounding box values during training and testing.

[xiankgx]

During training, we fit in batches of images of the same dimensions for training, 512 in code by default. During training the position of activated pixels to the rotated box boundaries is limited to the range [0, 512] due to the use of a sigmoid activation function. However, during testing, the model input is not restricted to the size of images used during training, instead, only resized to be divisible by 32. I'm wondering what's the effect of this when the test image dimensions are very different than the training image dimensions. Do you think it's better to squashed resized images to 512 during testing? @SakuraRiven

[SakuraRiven]

I guess "dimensions" is actually "scale" ? The scales in ICDAR2015 train/test are similar so we could directly inference. In fact, we can adjust the train scale and test scale to align them for better performance.