Gaussian Vector: An Efficient Solution for Facial Landmark Detection

Abstract

Significant progress has been made in facial landmark detection with the development of Convolutional Neural Networks. The widely-used algorithms can be classified into coordinate regression methods and heatmap based methods. However, the former loses spatial information, resulting in poor performance while the latter suffers from large output size or high post-processing complexity. This paper proposes a new solution, Gaussian Vector, to preserve the spatial information as well as reduce the output size and simplify the post-processing. Our method provides novel vector supervision and introduces Band Pooling Module to convert heatmap into a pair of vectors for each landmark. This is a plug-and-play component which is simple and effective. Moreover, Beyond Box Strategy is proposed to handle the landmarks out of the face bounding box. We evaluate our method on 300W, COFW, WFLW and JD-landmark. That the results significantly surpass previous works demonstrates the effectiveness of our approach.

LOTR: Face Landmark Localization Using Localization Transformer

“…• We demonstrate that the proposed LOTR framework detects facial landmarks accurately. Experimental results indicate the superiority of the proposed LOTR over other algorithms, such as two recent heatmap-based methods (Earp et al, 2021;Xiong et al, 2020), on the 1st JDlandmark localization challenge leaderboard .…”

“…Recently, heatmap-based approaches (e.g. Earp et al, 2021;Kowalski et al, 2017;Mahpod et al, 2018;Xiong et al, 2020) have extensively been used for face landmark localization tasks as they better utilize spatial information to boost the performance compared with coordinate regression methods. These methods predict spatial probability maps wherein each pixel is associated with the likelihood of the presence of a landmark location.…”

“…However, they rely on generating large heatmaps, increasing the post-processing complexity. Xiong et al (2020) propose vectorization of predicted heatmaps and supervising training using vector labels obtained from a quasi-Gaussian distribution. This process helps reduce the post-processing complexity, in addition to their proposed method ranking second on the JD-landmark challenge .…”

“…This process helps reduce the post-processing complexity, in addition to their proposed method ranking second on the JD-landmark challenge . Earp et al (2021) propose a heatmap regression approach that makes use of a bag of tricks and sub-pixel inference to surpass the performance of Xiong et al (2020) on the JDlandmark challenge leaderboard . However, Xiong et al (2020) and Earp et al (2021) still rely on an external post-processing step to achieve the final landmark locations.…”

“…Earp et al (2021) propose a heatmap regression approach that makes use of a bag of tricks and sub-pixel inference to surpass the performance of Xiong et al (2020) on the JDlandmark challenge leaderboard . However, Xiong et al (2020) and Earp et al (2021) still rely on an external post-processing step to achieve the final landmark locations. This additional post-processing step increases computational complexity and thereby results in an increase in inference time (Earp et al, 2021) and, therefore, tackling this issue is the primary motivation behind this paper.…”