Getting the point cloud with a custom wrapper

[tum-vision]

Getting the point cloud with a custom wrapper

[JakobEngel]

Hmm, I cannnot see anything in the code that looks wrong - however the depthmap looks very strange.
Maybe that's the initial, random depth map? It is quite normal that the first couple of keyframes are very noisy due to the random initialization - which should become better after some seconds of video.

[v0n0]

If so, how would I know when each depth map pixel has enough accuracy I can triangulate it and display it? I thought a keyframe would have solid hypothesis for each valid pixel, but maybe I misunderstood.

[JakobEngel]

in this context, accuracy is always relative:
It pixels in a keyframe are as accurate as they are going to get with respect to that frame.
However, if the camera moves a long distance, tracking error accumulates. If then there is a loop-closure, this accumulated error is removed - in the process of which the pose of keyframes, and with them the pose of the corresponding points, can change drastically. This happens e.g. at 2:48 in the youtube video ( http://youtu.be/GnuQzP3gty4 ): Large parts of the pointcloud are moved significantly.
Now of course this is an extreme example, but on a small scale this happens all the time when new loop-closures are found (see e.g. the foodcourt sequence: http://youtu.be/aBVXfqumTXc). You could do something like only update points if their position changes by more than a threshold, however that's not straightforward to compute efficiently anymore.

[v0n0]

Do you think a depth map fusion approach could be used then, when 3 or more keyframes agree on the same "candidate" point depth, then create the point in the visualization?

The other thing I don't understand is, if the inverse depth is calculated just for high gradient image areas, how comes almost all of the depth map is marked as valid and has covariance greater than zero?

Thanks

[v0n0]

The example that comes to mind is figure 6 "Variational inpainting" of the latest paper of yours, how did you paint the full resolution depth?

[amiltonwong]

@v0n0 ,
Where's the figure 6 "Variational inpainting"? Where can I find that paper?

[v0n0]

You can find all of the papers here: https://vision.in.tum.de/research/lsdslam?redirect=1

[romulogcerqueira]

@v0n0, are you using meshlab as your custom point cloud viewer?

[tsingjinyun]

@romulogcerqueira ,it's meshlab