TrackCam: 3D-aware Tracking Shots from Consumer Video

摘要

Panning and tracking shots are popular photography techniques inrnwhich the camera tracks a moving object and keeps it at the samernposition, resulting in an image where the moving foreground isrnsharp but the background is blurred accordingly, creating an artisticrnillustration of the foreground motion. Such shots however are hardrnto capture even for professionals, especially when the foregroundrnmotion is complex (e.g., non-linear motion trajectories).rnIn this work we propose a system to generate realistic, 3D-awarerntracking shots from consumer videos. We show how computer visionrntechniques such as segmentation and structure-from-motionrncan be used to lower the barrier and help novice users create highrnquality tracking shots that are physically plausible. We also introducerna pseudo 3D approach for relative depth estimation to avoidrnexpensive 3D reconstruction for improved robustness and a widerrnapplication range. We validate our system through extensive quantitativernand qualitative evaluations.