Abstract: As early as May this year, Alibaba Cloud has launched a short video solution. Recently, Alibaba Cloud once again took the lead in subverting the industry. It launched the first free face recognition SDK in the industry. Combined with its original short video capabilities, it has greatly lowered the entry threshold for face recognition + AR special effects + short video.
What you know about the face recognition industry
When it comes to face recognition, everyone will think of commercial payment, identity recognition, advertising, human-computer interaction, system public security and many other life scenarios. Although research on this technology began in the 1960s, the subject has only become more active in recent years. Taking the more common scene in daily life - social interaction as an example, through the support of face recognition, AR special effects, and moving picture elements, it can help users break social barriers and express themselves and share more three-dimensionally, personalized and interesting. Bit of life.
Most of the face recognition SDKs already on the market are expensive, often costing hundreds of thousands, and some manufacturers claim to be free, but they are actually based on network API calls, which are not suitable for recording short videos on the mobile terminal. in the social scene. This has left many teams who want to add short video AR special effects to the APP beyond the reach.
Alibaba Cloud and Hand Tao launch a free SDK for face recognition
Through the integration of group resources, Alibaba Cloud deeply integrates the face recognition SDK developed by the Hand Taoist team and the Alibaba Cloud short video SDK, and truly achieves real-time detection, recognition, and tracking that does not depend on network APIs.
At present, the professional version of the short video SDK on the official website already has the face recognition function developed by Taotao, which realizes the complete experience of face recognition + AR dynamic stickers, combined with the original short video collection, import, crop, edit, and synthesis , ultra-fast upload, media asset management, video transcoding, distribution acceleration, playback and other full-link capabilities, Alibaba Cloud can provide entrepreneurs with a one-stop solution, allowing each APP to easily realize new short video AR gameplay.
Introduction of key technologies and speed measurement of algorithm performance
Let's take a look at the specific application scenarios and technologies. After entering the shooting screen on the client, users can choose and match personalized materials such as dynamic stickers to realize AR special effects, so that short videos can be different from sci-fi, cute, and spoof. Effect. It mainly involves core technologies such as face detection, key point positioning, and tracking.
First, face detection is used to locate faces in videos. Simultaneous detection of multiple faces and handling of complex situations such as multi-angle and partial occlusion of faces are also properly handled in this step, so that faces can be found quickly and accurately. .
第二， 人脸的关键点定位，则是用于已知人脸所在位置的基础上，自动标注人脸的轮廓、五官位置，比如眼睛、鼻子、嘴巴、眉毛、耳朵等关键位置。阿里云提供人脸识别关键点个数多达68个，可以更准确的追踪五官，保证用户的体验。
第三， AR特效美化，根据已知的关键点位置，搭配上用户所选的动态贴纸，并根据捕捉不同的面部动作来变换AR特效，达到真实互动。
第四， 人脸追踪，视频是动态而非静止的，当用户脸部移动、转动时，阿里云SDK可以实现对关键点的追踪，可识别姿态范围为yaw±60°，pitch±45°，roll±45°，精准的捕捉动作，持续追踪动态贴纸和AR特效。
阿里云人脸识别SDK具有准确度高的特性，通过68个关键点检测和以上技术，实现平均错误率低于 5%，出现“对不上”这种尴尬场面的概率极低。据悉，阿里云未来也会推出商业版人脸识别的高级功能，满足更高级客户的需求。
在性能方面，阿里云人脸识别算法和其它厂商算法在测速上的区别如下：
注：以上测试480p的最小人脸尺寸为48*48（px）；720p的最小人脸尺寸为72*72
从上表可以看出，本人脸识别算法在同样机型、同等测试对象的条件下，测速表现大幅优于业内友商。经过阿里云集团手淘亿级日活跃用户产品的考验，性能方面毋庸置疑。由于该算法也应用于手淘相关业务之中，所以后续的迭代、维护都会有强有力的保障。
阿里云人脸识别SDK的免费开放，给短视频行业带来了无限的可能性。基于阿里云，创业者和用户们都有了更多新鲜的玩法，创新机遇随之而来，希望整个行业能产生更多元、更深入的探索。
原文链接：https://yq.aliyun.com/articles/216752?spm=5176.100244.teamhomeleft.1.dLHtjZ