The Design And Implementation Of Generic Object Detection Module In Magic Camera

Short video clips are now very popular in the world.The composing module is the most important part of a UGC short video clips app,and the effects in the compose module can make it easier for people to create a short video clip.In the current market,based on fast face detection algorithm,most short video clips apps support effects such as 2D stickers,3D stickers,face deformation,etc.In order to develop more interesting effects,new algorithms such as gesture recognition,body recognition and SLAM have also been applied to video effects.However,generic object recognition has not been used to create effects so far.The Magic Camera,which depends on generic object detection algorithm to add effects,is the composing module of Hippocampus,which is also a short video clips app.The Magic Camera has four modules,including the detection module,effect module,render module and business module.This thesis will make a detailed introduction of the detecting module and a brief introduction of the other modules.The detecting module depends on the generic object recognition algorithm and the object tracking algorithm.The main work of this thesis is to design strategies to combine these two algorithms and realize a stable and fast algorithm on Phone without changing the algorithms.The strategy of the module is to detect objects with low frequency,select the appropriate objects from the result as the input of the tracking module,and use the detecting result to adjust the tracking result.At last,the Magic Camera integrated with the detection module has been designed and realized.The Hippocampus has been released to App Store and become the only app which uses generic object detection algorithm to add effects on videos.