So I’ve heard of object detection architectures like Faster R-CNN and image segmentation architectures like Fully Convolutional Networks (FCN).
My understanding is that object detectors predict multiple bounding boxes (sometimes many different classes) while segmentation networks mainly try to predict a foreground region of interest from the background. Can anyone elaborate on more differences?
Also, as far as I know, these are both unsupervised methods which don’t need the ground truth labels or I am mistaken? I know that the R-CNN architectures are tested on PASCAL VOC and COCO for benchmark accuracy but for applications purposes does my dataset need to have ground-truth labeling?
They need labels. Once you’ve finished lesson 14 you’ll have seen how to implement segmentation, so hopefully it’ll be clear then.
In the absence of labeled medical data do you think a pre-trained segmentation network like FC-DenseNet or Mask R-CNN could be used to bootstrap binary masks (simple shapes like cells) like transfer learning but for segmentation?
If so, which labeled segmentation dataset do you think would transfer to circular cell membrane shapes best?
U-net segmentation works best on medical data. But you can try Tiramisu or mask r-cnn too.
My understanding is that U-net is a FCN with skip connections while Tiramisu is FC-Densenet. Is there anything specific to U-net that make it good for medical segmentation?