Fast Object Annotation in Point Clouds Aided by 3D Reconstruction

Supervised learning relies heavily on labeled datasets, but data annotation is time-consuming and tedious work. To speed up annotation, in this paper, we propose a fast data annotation method for RGB-D samples collected in static indoor environments. The method first uses 3D SLAM technology to get a 3D map (the result of 3D reconstruction) of the environment and the 6Dof collection points of the RGB-D samples. In this process, multiple object instances appearing in consecutive RGB-D samples are fused into one object, as long as they belong to the same object in the physical world. Then we manually annotate objects in the 3D map to get global object labels that are represented in the world coordinate system. In this way, we only need to annotate a physical object in the 3D map once. Subsequently, a post-processing module converts every global label to many labels which are compatible with RGB-D samples. For an object existing in
$n$
RGB-D samples at the same time, after the once annotation in the 3D map, our method could automatically generate
$n$
labels for the
$n$
RGB-D samples. Therefore, our method has the efficiency that annotating once is equivalent to annotating all.