ashnair1 / mask-rcnn-for-off-nadir-building-detection Goto Github PK

Hi there! It looks like requirements.txt is missing in the repository.

am using maskrcnn for the detection of oil slicks on the sea surface using radar images.
In fact, I trained Maskrcnn on a dataset of SAR images(tif). Everything works fine for the training. However, I have some issues to run the prediction on big images. When I give as input small images (2000,2000), it works. However, when I try bigger sized images, for instance, (8000,5000), it generates the error below:
Resource exhausted: OOM when allocating tensor with shape[1,64,7264,7264] .

Could you please give me more insight on the above issues?

I am trying to solve this by dividing the images into crops and running the prediction on every crop, then reconstructing the whole result of the image. The problem is that I don't know how to reconstruct the hole result:
for each crop I get the dictionary of the below arrays.
The dict contains -masks: [H, W, N] instance binary masks.
-rois: [N, (y1, x1, y2, x2)] detection bounding boxes
-class_ids: [N] int class IDs
-scores: [N] float probability scores for the class IDs
I manage to reconstruct the mask as I know the position of the crop in the big image.
However, for the rest of the dictionary elements, I can't figure out how to group them properly in one array each with respect to the original large image.
I thank you a lot in advance for your collaboration.