At the recently finished Grand Challenge of 106-p Facial Landmark Localization, one of the grand challenges organized for the IEEE International Conference on Multimedia & Expo (ICME) 2019, Baidu’s visual tech team has reaped the first place, beating the other more than 20 competing teams.

▲ The ranking list of ICME 2019 Grand Challenge of 106-p Facial Landmark Localization

             ▲Cumulative Error Distribution of the top 3 competing teams 

Organized by IEEE’s Computer Society, Circuits and Systems Society, Signal Processing Society and Communications Society, IEEE ICMEis a flagship multimedia conference.It has been held for 18 consecutive years, covering topics such as text analysis and the processing, transmitting, analysis and application of multimedia data including graphs & images, videos, audios and more. 

The Grand Challenge of 106-p Facial Landmark Localization is one of the four challenges in ICME 2019. 

Facial Landmark Localization, as the key step of many applications such as face recognition, facial expression estimation and facial reconstruction, has witnessed many developments in recent years. This challenge was mainly focused on testing the robustness and accuracy of algorithms in complex scenarios when gesture, extreme expression or occlusion occurs.

In this challenge, Baidu-Vis team from Baidu’s visual technology department deployed Baidu’s AutoML to conduct Neural Architecture Search and introduced scattered density supervisionand weighted heatmap supervisionto the training model. At the same time, Baidu-Vis team used consistent estimation algorithm to revise prediction, and eventually improved the robustness and accuracy of algorithmsignificantly. All these approaches helped Baidu win the first prize of the challenge.