Learning to Segment Every Thing

[NHZlX]

https://arxiv.org/abs/1711.10370 (Submitted on 28 Nov 2017)

[NHZlX]

[NHZlX]

Train

• A data with bbox and mask info
• B data just with bbox info
• train the detection branch using A data
• train the mask branch using B data with detection branch fixed
• train the two branch end-2-end(the bp of Wseg don't affect the Wdet).

Proposal partially supervised learning(a novel transfer learning), transfer the learned detection branch info to mask branch using weight transfer function.

Use class-agnostic mask prediction of mask rcnn as base-line.