. What can semantic segmentation not distinguish, that instance segmentation adds?
Semantic segmentation only assigns a class per pixel; it has no notion of separate identity, so two touching objects of the same class merge into one labeled region.
. How does this lesson turn Lesson 42's segmentation network into an instance segmentation network?
The network keeps its original semantic head and gains a second head trained to regress a per-pixel offset to the instance center; clustering the resulting votes (Hough-style) recovers individual instances.
. In the toy mask head experiment, why can the mask head reach high per-instance IoU even though every test crop contains a visible neighboring circle?
Given a box already centered on one instance, the mask head's job narrows to a local same-box question, which is why Mask R-CNN's real mask branch can be small and fast — a good box already does most of the work.
. What is the one new piece Mask R-CNN adds on top of Faster R-CNN (Lesson 41), and what other change does it require?
Mask R-CNN keeps Faster R-CNN's RPN, classifier, and box regression unchanged, adds a mask branch per box, and swaps RoIPool for RoIAlign because coordinate quantization that barely hurts classification visibly misaligns pixel-accurate masks.