Sogou Excels in Facial Landmark Localization Challenge

The Chinese computer vision developer Sogou has bested the competition in the 106-Point Facial Landmark Localization Grand Challenge at this year’s International Conference on Pattern Recognition (ICPR). The conference is now in its 25^th year, and is hosted by the International Association for Pattern Recognition (IADR).

This is not the first time that Sogou has come out ahead in an international computer vision competition. The company took the top prize in the Autonomous Driving Challenge at the CVPR Workshop in 2018, and then posted the top score in the MegaFace Million-Scale Face Recognition Challenge. The company’s algorithm was 99.939 percent accurate in the latter case.

Sogou used an enhanced HRNet-based convolution network and a Pose-based Data Balancing (PDB) strategy to defeat its rivals in the ICPR challenge. The former made the company’s model more efficient and boosted the accuracy of its algorithm, while the latter helped predict the position of different facial landmarks to solve problems with unbalanced data. Together, those technologies led the company’s AI Interaction Division to the top of the leaderboards in both the validation phase and the overall final evaluation.

Sogou noted that facial landmark localization plays a key role in many face-based applications, including facial recognition, pose estimation, and image synthesis. The Grand Challenge tests an algorithm’s ability to generalize with a dataset that has large variations in factors like pose, expression, and occlusion. It also places limits on the weights that competitors can use to make their computational models more efficient.

Carnegie Mellon researchers recently developed a solution that can create 3D reconstructions of people’s faces using video footage captured with a smartphone camera. CyberLink and Google have also performed well in previous iterations of the MegaFace Challenge.

–

November 4, 2020 – by Eric Weiss