Among the well-established methods for facial landmark detection is the family of Constrained Local Models (CLMs). An important part of CLM landmark alignment pipeline are the local detectors that estimate the alignment probability of each individual landmark over the facial region. In this paper we present Deep Constrained Local Model (DCLM) algorithm and the novel Dense-Projection Network (DPN) as a local detector. DPN is a deep neural network that consists of two important layers: Template Projection layer and Dense Aggregate layer. In Template Projection layer, patches of facial region are mapped to a higher dimensional space allowing the pose and rotation variations to be captured accurately. In Dense Aggregate layer an ensemble of experts is simulated within one network to make the landmark localization task more robust. In our extensive set of experiments we show that DPNs outperform previously proposed local detectors. Furthermore, we demonstrate that our proposed DCLM algorithm is state-of-the-art in facial landmark detection. We significantly outperform competitive baselines, that use both CLM-based and cascaded regression approaches, by a large margin on three publicly-available datasets for image and video landmark detection.
from cs.AI updates on arXiv.org http://ift.tt/2gA6E08
via IFTTT
No comments:
Post a Comment