When using rgbdslam to generate point clouds of a room interior using a kinect, the camera needs to follow a complex path around the room. Even when doing this, there are significant holes in the resulting model, and many areas are scanned several times. To illustrate the point, a simple helical scan from the center of a room will produce many shadows (holes) in the resulting point cloud, due to the objects in the room.
Getting different viewpoints from around the room is required to get a full point cloud model of the room. In order to fill in these holes, you need to position the camera at various viewpoints to get the missing points in the cloud.
It looks there is a need for a feature or separate tool that would allow for flipping between the individual pcd nodes and selecting one that have neighboring areas with missing coverage, so that IPC can continue on new frames that should be added to the specific region. Without this, the only other option is to pan back through the model to get to the areas where these holes exist. This re-coverage over existing points usually causes distortions in the cloud due to minor miss alignments in registration.
It also seems like assembly of individual PCD nodes that have been saved for later assembly, could be used to build up a complete model, where wholes can be identified and then specific scans are taken to fill these areas. This again brings up the need for another tool, or feature.
Does anyone know of a good example of point clouds that are relativity complete models of interior volumes, that use the kinect?
Walt
Originally posted by walt on ROS Answers with karma: 16 on 2011-10-14
Post score: 0
Original comments
Comment by Felix Endres on 2011-10-17:
Except for the last request, I do not understand what it is you are asking for. You could browse through the sequences of the RGB-D Benchmark at http://cvpr.in.tum.de/data/datasets/rgbd-dataset - Maybe there is a "complete" dataset.