For multi instance segmentation tasks like cell segmentation, would using a mask or bounding box be useful?

Question

While dealing with image segmentation tasks using models like the UNet family(Unet++, double UNet, ResUNet..), most of the tasks were to find one specific instance, meaning the mask was a single object like detecting the left heart from the Ultra Sound medical images.

However this time, I am currently handling a cell segmentation task with a lot of instances to segment as a mask. (One class though)

I am curious if using a RoI generating model like MaskRCNN would be better this case than a UNet like network.

Also, is it better to use the pretrained network like pytorch segmentation models? (I used to make and train all my models from scratch.)

Thank you.

Answer 1

Your problem definition does sound more like instance segmentation than semantic segmentation. Therefore, it is more "natural" to use architectures designed for that specific task.
Furthermore, in general, it is easier to re-use existing models/code rather than re-implementing them from scratch.