We want to simplify the production flow of a special effect that used to be operated on Pr and AE, which is tracking one character and always fixing it in the middle of the screen. This special effect usually applies in motion videos, such as skiing video, dance video, surfing video, etc. To make that effect, editors need to set many keyframes and adjust screen size all the time in Pr, which need much knowledge about the editing software and spend much time on editing. Our program can let users input an original video, then directly generates an edited video with that special effect.
We used pre-trained data of YOLOv3 and TensorFlow.
Unprocessed video:
Link
We trained an object detector with TensorFlow Object Detection module and YOLOv3 sprarately to identify and locate objects in an image or video.
After we get the location of the object along with its label, we proportionally resized each frame of a video so that the object is centered.
Github repo:
Link
Processed video:
Link