As face landmark localization is a vital module of various computer vision tasks, the Sertis Computer Vision presented a state-of-the-art face landmark localization model that improved performance on the face recognition task.
Now, we are in the industrial revolution era where we are able to develop the ‘eyes’ for computers or what we call Computer Vision. This technological advancement has led to the emergence of further technologies that bring many benefits to humans, such as face recognition technology to be used in security systems, or the use of object detection to define objects and human pose estimation for sports analytics and people tracking.
At Sertis, we have foreseen the benefits that can be extended from such technology to more advanced technology. So we have set up a team of researchers of Sertis Computer Vision who have worked on researching and developing Computer Vision continuously, apart from Sertis Face Recognition system.
Sertis AI research team, including Samuel W. F. Earp, Aubin Samacoits, Sanjana Jain, Pavit Noinongyao and Siwa Boonpunmongkol researched and developed a method for face landmark localization using heatmap regression approach, with several different tricks and techniques for a more efficient and accurate model.
Face landmark localization is an important process of the detection and localization of certain key points on the face, which has become a widely used technology that benefits to many industries for several purposes such as for a security system or for entertainment. In this research, Sertis AI research team presents the heatmap regression approach with each model consisting of a MobileNetV2 backbone followed by several upscaling layers, with different tricks to optimize both performance and inference cost.
In this experiment, the team has used five naïve face landmarks from a publicly available face detector to position and align the face instead of using the traditional Bounding Box methods. Moreover, they added random rotation, displacement, and scaling, after face alignment and observed that the model is more sensitive to the face position than orientation. The team showed that it is possible to reduce the upscaling complexity by using a mixture of deconvolution and pixel-shuffle layers without impeding localization performance. They also incorporated a sub-pixel inference module to obtain more accurate landmark positions.
With all these tricks and techniques, the team was able to develop a new state-of-the-art face landmark localization model that tested the effect and evaluated the performance on face recognition using a publicly available model and benchmarks.
In the 2nd Grand Challenge of 106-Point Facial Landmark Localization competition, Sertis’s face landmark localization model ranked second on the validation set test. This success has marked another achievement for Sertis to push ourselves towards the further development of useful technology and being part of the industry to support artificial intelligence technology advancement in the long run.