Human Part Semantic Segmentation Using Custom-CDGNet Network

ADITI VERMA, Vivek Tiwari, Mayank Lovanshi, Rahul Shrivastava

Abstract:

Human body part segmentation is a semantic segmentation of human
images task that entails labelling pixels in an image into their respective classes.
The human body is composed of hierarchical structures in which each body part
in the image has a particular individual location. Considering this knowledge, the
sample class distribution technique was developed by collecting and applying the
primary human parsing labels in vertical and horizontal dimensions. The proposed
network exploits the underlying position distribution of the classes to make precise
predictions with the help of these classes. We produce a distinct spatial guidance
map by combining these guided features. This guidance map is then superimposed
on our backbone network. Extensive experiments were executed on a large data set,
i.e. LIP, and evaluation was done using the mean IOU and MSE-loss metrics. The
proposed deep learning- based model surpasses the baseline model and adjacent
state-of-the-art techniques with a 2.3% hike in pixel accuracy and a 1.4% increase
in mean accuracy