Why is validation step slower than training?

[zungam]

I am training on a dataset of 19 images. Also, I am doing validation on 17 images. Time training 19 images = 30 seconds. Time validating 17 images = 200 seconds. Why is this happening? Domains is completly similar.

Each image has about 100 masks. Yes, its alot. Maskrcnn is probably not made to handle so many, but I am trying to fix just that.

I see that during validation, it spend alot of time computing overlaps, its a bottleneck. But I dont understand why during training it does not consider this to be a bottle neck, why?

[artisvirat]

Hey Zungam, Can you please explain in detail how you converted your dataset into coco format? Much Appreciated Thanks

[zungam]

I recomend following this tutorial https://engineering.matterport.com/splash-of-color-instance-segmentation-with-mask-r-cnn-and-tensorflow-7c761e238b46