Instance Segmentation

Combination of Semantic Segmentation and Object Detection

Detect all objects in the image, and identify the pixels that belong to each

  • Only things not stuff
    • Objects that can be separated into object instances
Approach

  1. perform object detection
  2. predict a segmentation mask for each object

Mask R-CNN

Adding onto Object Detection#Faster R-CNN

Attach a separate branch for segmentation

Pasted image 20241205134806.png
Pasted image 20241205134834.png

Example Training Targets

A pair of boundary box, category and segmentation mask
Pasted image 20241205135000.png

Panoptic Segmentation

Label all pixels in the image

Keypoint Estimation

Pasted image 20241205135209.png
Attach another branch on Mask R-CNN
Pasted image 20241205135233.png

We can attach different branches to achieve different final results