HoVer-Net Review

Problem Statement from HoVer-Net

• Mask R-CNN has great performance in clustered image segmentation, becuase Region proposal strategy selects each object and the segmentation process are performed at the each selected regions.
• However, HoVer-Net insists that Mask R-CNN has difficulty in merging instance predictions.
  • Mask R-CNN uses Region Proposal, and Region Proposal selects up 2,000 Region of Interest(ROI), and the segmentation is done in the each selected ROI.
  • each ROI is already selected considering each single ROI, so it doesn’t have to be separating cells furtherly.

#### < Region Proposal > As it can be seen in the 2. Extract region proposals, the region proposals selects ROI.

• HoVer-Net says nucleous locating on the boundary of the ROI (yellow boxes in the image above) could be split into multiple parts and would be located in multiple tiles.
  • Each two tiles which are green boxes has subsegments of nucleous.
  • It requires that left half and right half of the nucleous has te be labeled as one single label, and it takes more efforts to make sure they have identical

Solution that HoVer-Net Propses

HoVer-Net uses Overlapping Tile Strategy that U-Net firstly introduced to make segmentation output with seamless.

• Using Overlap Tile Strategy, HoVer-Net doesn’t have to consider boundary and also doesn’t have to think about the split of the nucleus that requires consistent label in splitted ones.