Abstract

We present PixTrack, a vision based object pose tracking framework usingnovel view synthesis and deep feature-metric alignment. Our evaluationsdemonstrate that our method produces highly accurate, robust, and jitter-free6DoF pose estimates of objects in RGB images without the need of any dataannotation or trajectory smoothing. Our method is also computationallyefficient making it easy to have multi-object tracking with no alteration toour method and just using CPU multiprocessing.

 

Quick Read (beta)