3

I've been exploring 3D scanning and reconstruction using Google's project Tango.

So far, some apps I've tried like Project Tango Constructor and Voxxlr do a nice job over short time-spans (I would be happy to get recommendations for other potential scanning apps). The problem is, regardless of the app, if I run it long enough the scans accumulate so much drift that eventually everything is misaligned and ruined.

High chance of drift also occurs whenever I point the device over a featureless space like a blank wall, or when I point the cameras upward to scan ceilings. The device gets disoriented temporarily, thereby destroying the alignment of future scans. Whatever the case, getting the device to know where it is and what it is pointing at is a problem for me.

I know that some of the 3D scanning apps use Area Learning to some extent, since these apps ask me for permission to allow area learning upon startup of the app. I presume that this is to help localize the device and stabilize its pose (please correct me if this is inaccurate).

From the apps I've tried, I have never been given an option to load my own ADF. My understanding is that loading in a carefully learned feature-rich ADF helps to better anchor the device pose. Is there a reason for this dearth of apps that allow users to load in their homemade ADFs? Is it hard/impossible to do? Are current apps already optimally leveraging on area learning to localize, and is it the case that no self-recorded ADF I provide could ever do better?

I would appreciate any pointers/instruction on this topic - the method and efficacy of using ADFs in 3D scanning and reconstruction is not clearly documented. Ultimately, I'm looking for a way to use the Tango to make high quality 3D scans. If ADFs are not needed in the picture, that's fine. If the answer is that I'm endeavoring on an impossible task, I'd like to know as well.

If off-the-shelf solutions are not yet available, I am also willing to try to process the point cloud myself, though I have a feeling its probably much easier said than done.

Eugene Ang
  • 31
  • 1

1 Answers1

3

Unfortunately, Tango doesn't have any application could do this at the moment, you will need to develop you own application for this. Just in case you wonder how to do this in code, here are the steps: First, the learning mode of the application should be on. When we turn the learning mode on, the system will start to record an ADF, which allows the application to see a existing area that it has been to. For each point cloud we have saved, we should save the timestamp that associated with points as well.

After walking around and collecting the points, we will need to call the TangoService_saveAreaDescription function from the API. This step does some optimization on each key poses saved in the system. After done saving, we need to use the timestamp saved with point cloud to query to optimized pose again, to do that, we use the functionTangoService_getPoseAtTime. After this step, you will see the point cloud set to the right transformation, and the points will be overlapped together.

Just as a recap of the steps:

  1. Turn on learning mode in Tango config.

  2. Walk around, save point cloud along with the timestamp associated with the point cloud.

  3. Call save TangoService_saveAreaDescription function.

  4. After saving is done, call TangoServcie_getPoseAtTime to query the optimized pose based on the timestamp saved with the point cloud.

xuguo
  • 1,816
  • 1
  • 11
  • 15
  • Hi Jason, thank you for getting back to me. Should I use the Java or C API? Your recommendation implies that C is the way to go. Is there a difference? Just to clarify, after I save the area description file, when I call getPoseAtTime() for each point cloud saved, is there something I need to do to the TangoPose result returned by getPoseAtTime() to correct the drift of each point cloud, or will the drift already be corrected after the function call? – Eugene Ang Mar 23 '16 at 16:47
  • you could use c, java or unity. They are all supported. The drift correction (loop closure) happens when the `TangoService_saveAreaDescription` is called. So the query with `TangoServcie_getPoseAtTime` after saving will return back the corrected pose. – xuguo Mar 23 '16 at 17:28