DeepTemplates: Object Segmentation Using Shape Templates

Nikhar Maheshwari, Gaurav Ramola, Sudha Velusamy, Raviprasad Mohan Kini

Abstract: Object segmentation is a crucial component for many vision applications like object editing and augmented reality. Traditional pixel-wise segmentation techniques result in irregularities around object boundaries. However, applications like lip makeup require smooth boundaries that resemble a typical lip “template-shape”. We propose an encoder-decoder architecture that is explicitly conditioned to utilize the underlying template-shape of an object for segmentation. The decoder was trained separately to generate a template-shaped segment obtained from landmark points of an object. The encoder is then trained to predict these landmarks using Euclidean loss. Finally, we jointly train the encoder and decoder by incorporating the decoder’s segmentation loss to refine the landmarks, which conditions the network to produce template-shaped object segments. The performance of the proposed method was evaluated with mIOU and f-score measures on the HELEN data set for lip segmentation. We observed perceptually superior segments with smooth object boundaries when compared to state-of-the-art techniques.

External IDs:dblp:conf/cvip/MaheshwariRVK22