Efficient interfaces for accurate annotation of 3D point clouds

Abstract:

Collecting massive 3D scans of real world environments has become a common practice
for many private companies and government agencies. This data represents the
real world accurately and densely enough to provide impressive visualizations. However,
these scans are merely points. To truly tap into the potential that such a precise
digital depiction of the world offers, these scans need to be annotated in terms of objects
these points represent. Manually annotating this data is very costly. Existing
machine-aided methods report high accuracies for object localization and segmentation.
However, the central task of annotation, proper label assignment, is still a
challenging task for these approaches.
The goal of this work is to design an interface that streamlines the process of
labeling objects in large 3D point clouds. Since automatic methods are inaccurate
and manual annotation is tedious, this work assumes the necessity of every object’s
label being verified by the annotator, yet puts the effort required from the user to
accomplish the task without loss of accuracy at the center stage.
Inspired by work done in related fields of image, video and text annotation, techniques
used in machine learning, and perceptual psychology, this work offers and
evaluates three interaction models and annotation interfaces for object labeling in 3D
LiDAR scans. The first interface leaves the control over the annotation session in
the user’s hands and offers additional tools, such as online prediction updates, group
selection and filtering, to increase the throughput of the information flow from the
user to the machine. In the second interface, the non-essential yet time consuming
tasks (e.g., scene navigation, selection decisions) are relayed onto a machine by employing
an active learning approach to diminish user fatigue and distraction by these
non-essential tasks. Finally, a third hybrid approach is proposed—a group active
interface. It queries objects in groups that are easy to understand and label together
thus aiming to achieve advantages offered by either of the first two interfaces. Emiii
pirical evaluation of this approach indicates an improvement by a factor of 1.7 in
annotation time compared to other methods discussed without loss in accuracy.