Image size for training Mask Rcnn

[vsd550]

My image dataset has images of different sizes ranging from 600 x 600 to 3000 x 3000. Does the Mask Rcnn take a fixed size input image i.e. do I need any preprocessing before passing the images for training. I have made the coco-like json file for training with my dataset.
Thanks in advance

[roytseng-tw]

The preprocessing about image size is controlled by the flags: TRAIN.SCALES, TRAIN.MAX_SIZE.
An image is resized to make the smaller side as in one of the scale randomly chosen from TRAIN.SCALES, if the resized longer side doesn't exceed TRAIN.MAX_SIZE. Otherwise, image is resized to make the longer side equals to TRAIN.MAX_SIZE.

[vsd550]

Thanks a lot