Paper NOTE: Semantic segmentation using adversarial Networks

Semantic Segmentation using adversarial Networks
2018-04-27 09:36:48

Abstract:

For the production of image modeling, the confrontation training has achieved very good results. In this paper, we propose a method of antagonistic training to train semantic segmentation model. In fact, this is the addition of a confrontation loss, that is: a CNN to determine the given graph is the result of segmentation? Or the GT? The motivation of this method is: it can detect and correct Higher-order inconsistencies between GT segmentation maps and the Ones produced by the segmentation net.

The contributing points of this article are:

1. The first attempt to introduce adversarial training into the semantic segmentation field;

2. This method guarantees: long-range spatial label contiguity, and does not add complexity when testing;

3. Improved performance on two datasets;

　　

The proposed approach:

　　Adversarial training :

The method of this article is to use two mixed loss function, the first item is: A multi-class cross-entropy term, which estimates the segmentation model to independently predict the category label for each pixel location.

Given an RGB image X, the partitioning model outputs a class probability plot (the class probability map) s (x);

The second item is based on an additional anti-convolution network.

　　

==============================================

　　

　　

　　

　　The Network Architecture:

　　

According to the above flowchart can be found, this article is to divide the result/GT two value diagram and the original image is multiplied, the results obtained, input into the confrontation network. The specifics are as follows:

　　

　　

======================

Reference:

1. Chainer Implementation:https://github.com/oyam/semantic-segmentation-using-adversarial-networks

2. Pytorch implementation:https://github.com/gzhermit/pytorch-gan4segmentation

Paper NOTE: Semantic segmentation using adversarial Networks