[Advanced] Semantic Segmentation-Starter Toolkit


#1

Hi folks,
I’m trying to get jumpstarted with implementing a Semantic Segmentation pipeline.

Please help me prepare a Starter toolkit for the same. Where should I start?
I think I’ll be using PyTorch.

Sanyam


#2

Where should I start?

There is a good paper published last year that provides a review on deep learning methods for semantic segmentation applied to various application areas. https://arxiv.org/abs/1704.06857

There’s more:

Semantic segmentation pipelines

Typical pipeline use a variant of Fully Convolutional Network (FCN) piped into a Conditional Random Fields (CRF). CRF is one possible and common approach to refine the output of a segmentation system and boost its ability to capture fine-grained details. FCN method basically is using sliding window while CRF seems to assign random variable for every pixel.

I recommend you read some of the top papers from the segmentation results in PASCAL VOC Challenge to get a general idea of the best methods currently.

http://host.robots.ox.ac.uk:8080/leaderboard/displaylb.php?challengeid=11&compid=6

Hope that helps.


#3

On it!
Thanks a lot! @cedric
Do you think it’ll be an easy implementation attempt? I haven’t ever tried implementing a paper end to end. I’ve always been scared of it, :sweat_smile: