Sogou wins facial landmark localization challenge at ICPR 2020
Chinese search engine company Sogou has taken first place ranking in a facial landmark localization challenge at the 25th International Conference on Pattern Recognition (ICPR 2020), demonstrating the company’s technological strength and position as an innovator in the computer vision space, according to a company announcement.
Facial landmark localization is important for a number of applications related to computer vision, such as face alignment for biometric recognition, facial pose estimation, and face image synthesis, Sogou says. The company argues in the press release that exploring cutting edge artificial intelligence technologies like facial recognition, voice recognition and natural language processing will bring the best user experience possible to its products and services in the future.
The ‘Grand Challenge of 106-Point Facial Landmark Localization’ tests the accuracy and robustness of the algorithm’s generalization ability. The challenge makes use of a dataset with images that vary widely by identity, pose, expression and occlusion, and also imposes strict limits of model weight for computational efficiency.
Sogou held the top spot in both the validation phase and the final evaluation, leading teams from a number of highly-regarded enterprises and research institutes. The company leveraged technologies including enhanced HRNet-based structure and group convolution networking to deliver the highest possible accuracy and efficiently reduce the complexity of its model for facial landmark localization. The team also adopted a Pose-based Data Balancing (PDB) strategy to resolve unbalanced data in various positions.
ICPR 2020, hosted by the International Association for Pattern Recognition (IADR), will be held as an online event from January 10 to 15, 2021.
The company also cites its biometric technology’s true acceptance rate (TAR) of 99.939 percent in the MegaFace Million-Scale Face Recognition Challenge in 2018 and its first prize in the CVPR 2018 Workshop on Autonomous Driving Challenge as evidence of its strong momentum in computer vision research and development.