3D Mapping

Hello,

This technique (essentially a moving version of photogrammetry) looks very interesting:

https://www.facebook.com/wkajohnson/videos/10157387450590335

Real footage top left is converted into 3d geometry and then textured to create a virtual version (bottom right).

At a guess this looks like it relies on good footage to make the calculations but does look like it is a semi automated process.

Interesting stuff but the facebook post is a bit light on detail, I'd like to know which softwares he is using? I've had a go with similar optical flow and structure from motion in the past in caves but found that the softwre got confused due to moving light sources. I guess things move on and he has some pretty good lighting

TheBitterEnd said:

Interesting stuff but the facebook post is a bit light on detail, I'd like to know which softwares he is using? I've had a go with similar optical flow and structure from motion in the past in caves but found that the softwre got confused due to moving light sources. I guess things move on and he has some pretty good lighting

Click to expand...

Would be interesting to see which software this person used, yes.

I had a go at doing something like this a couple of weekends ago at the Alderley Edge open weekend, using VisualSFM to create a point-cloud from a GoPro video (or rather extracted frames from a GoPro video).
From first look, it worked well where we had part of the mine lit-up, and it did pretty well in some other bits of the mine where I was relying on my own light sources - even though the lighting was changing. The difficulty I found was that in the bit where I was providing the lighting the GoPro stills showed some motion blurring - presumably because it was still quite dark. This meant that some of the features it was trying to match weren't very distinct. The software did get a bit confused at some points due to this and started creating a couple of separate models not linked by any common points.

I was using the mode of looking for common features on one photo against all other photos - and this starts to take a very long time after you've got a few hundred frames extracted. But you have too big a time gap between extracted frames on the video and you start getting lower number of matched points. I understand that it's possible to look only at n photos either side of the current one which should make the run times more sensible - but not sure how that would deal with a situation where your video frames came back to the same point - so want to match the features at the start and end of the video.

Presumably you'd know where your loop closure is? If you can identify the frame would you just say, 'Scan current frame +/- n frames and (known reference closure frame)' for matches.

Easier said than done I'm sure. My experience is practically nil on the computer vision front.

Hi folks - the main contact is one of the lead devs at Unity (makers of game engine software)....but this looks like his pet project

The way this works "looks" like some pretty clever image analysis software - it's not only tracking the camera movement but interpreting the movements into geometry.

Hopefully he will get back to the PM I've dropped him...