Mask_rcnn: Improvement - Faster Training of Mask R-CNN by Focusing on Instance Boundaries

[Paulito-7]

This is not an issue but rather a suggestion of improvement.

I saw the publication of a paper in September showing great improvements for Mask R-CNN performance called https://arxiv.org/abs/1809.07069. Besides accelerating the training process (29% faster according to the paper), it enables to have better edges and less false positves, and seems not to be complex to implement, starting from the implementation of Mask R-CNN.

Shall we add this as an improvement?

[fastlater]

@Paulito-7 I was reading the same paper a few days ago too. The idea sounds good. I hope someone adds this improvement.

[chengchu88]

+1

[JulienSiems]

Our implementation is available here: https://github.com/FlashTek/mask-rcnn-edge-agreement-loss
We only had to make minor adjustments to the code, mainly here for the edge agreement loss and here to register it as an additional loss.

[nicejava]

Our implementation is available here: https://github.com/FlashTek/mask-rcnn-edge-agreement-loss
We only had to make minor adjustments to the code, mainly here for the edge agreement loss and here to register it as an additional loss.

How to config model.py for high accuracy boundaries ? you have example @JulienSiems
Now i try follow this link https://github.com/FlashTek/mask-rcnn-edge-agreement-loss but have bad boundaries like this