0
$\begingroup$

I am trying to 3D reconstruct a room that has been freshly constructed but the walls have not yet been either plastered or painted/wallpapered. So far I have tried using mapping techniques(like rtab map) on KinectV2 but they don't work as these techniques rely on features to stitch point clouds.

I am currently looking at buying either Structure sensor or Google's Project Tango on Phab2Pro. Since, these are a little expensive (post shipping and customs) I want to be sure of a few things before I begin experimenting.

  1. Do these sensors use something other than features to register point clouds (the phone's accelerometer, for example)?
  2. Is one sensor better at it than the other at this job? If so, why?

If any one of you could somehow attach a point cloud or an image captured using these sensors, it would be of plenty help. Also, feel free to suggest better alternatives.

Thanks!

| improve this question | |
$\endgroup$
  • $\begingroup$ Would not be cheaper to just stick old newpapers on the walls for a moment, so providing a lot of points for Kinetic? $\endgroup$ – gilhad May 2 '17 at 14:45
  • $\begingroup$ If it is a featureless environment to the extent that pictures taken cannot be stiched together, why does it need to be scanned? Can it be simply modelled based on its geometry? $\endgroup$ – 50k4 May 3 '17 at 11:36
1
$\begingroup$

For Kinect Fusion, and mostly all other point cloud fusion algorithms, ICP is used for aligning the clouds and creating the mesh. In a feature-less scene, ICP does not work, as there is no easy way to calculate local minimum errors to align the clouds so the stitching is almost impossible over a larger space. There is additional research work available online for eg (https://graphics.stanford.edu/~mdfisher/papers/inertialNavigation.pdf) that use accelerometer data to improve the fusion process, but if there are no features to align, the accelerometer data might not be able to help much (From the paper: "We use the rigid transform estimated by inertial navigation to provide a significantly better initial guess for the ICP algorithm". ICP would still require features to align).

Just for reference, I'm working on Kinect in an environment that limits the point cloud generation to a very small region. Since the point cloud reduces and becomes sparse, i'm facing the same issue of ICP alignment, as the number of features are reduced.

An easier solution for your case would be to temporarily hang a few pictures or wall hanging accessories, as well as place a few plants or objects etc on the ground to create some features that help ICP align the clouds.

| improve this answer | |
$\endgroup$
  • $\begingroup$ I second this. You need features or you will lose all tracking ability. $\endgroup$ – MindS1 Oct 25 '18 at 13:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.