Masks are consuming huge memory

[vineel96]

Hi @akshitac8,
Masks are consuming huge memory as there are huge number of instances for some images (Ex. 1239*800*800 mask for a image).How did you handle this while implementing?
Also, could you give system details of which GPU you have used and memory of each GPU used for implementing Mask-RCNN as shown in ISAID research paper?

Thanks

[akshitac8]

Hello @vineel96
For ISAID paper, I used v100 GPU with 32GB memory.
Yes the mask will take up a good chunk of the memory but you can limit that based on how many instances you want to work on and how much your system can handle.

[vineel96]

Thanks @akshitac8 for the information.
How to filter instances? can we do it randomly? suppose if image has 2000 instances, if my gpu can handle upto 1000 instances, then which 1000 of 2000 instances should we take? because if we take first 1000 instances for training and if testing data has samples(say >100 images) with other 1000 instances then there is chance of decrease in test mAP right?