Single walls are constructed but not whole room

[prashantmore]

We are tying to create a 3D model of room.
We are able to generate 3D model of single wall at a time.
But when we are trying to create a 3D model of whole room we are not able to generate.
Also with captured images of 3 walls lot of black holes is seen in the model but the black holes is not seen in single wall captured.

I have captured images of 3 walls from office workstation.
There 3 walls are namely

1. Office branding wall.
2. Whiteboard wall.
3. Wooden board wall.

I have attached captured images.

Wherever we found black holes on wall, we stick cartoons drawn on papers and we minimized the black holes.

We also tried to capture images with different techniques which are mentioned below.
1. Keeping camera in center and capture all 3 walls.
2. Capture images by 90 degree of camera parallel to wall.
3. Capture images by 60 degree of camera parallel to wall.
4. Capture images by 45 degree of camera parallel to wall.
5. Capture images far away 10 feet from wall to take wide area.
6. Capture images far away 20 feet from wall to take more wide area.
7. Capture images by 90 degree of camera parallel to wall. Additionally captured 2 more images with every picture which includes inclined upside and inclined downside to get more overlaps.


We took care of motion blur problems by using tripod and Vibration Reduction turned ON.

We tried this experiment on different types of cameras.
1. 12 MP camera of smartphone. (With and without HDR)
2. 13 MP camera of smartphone.
3. 16 MP camera of smartphone.
4. 20.4 MP camera of Sony Cybershot.
5. 24.1 MP camera of Nikon D3200

Here is the google drive link of datatset :- https://drive.google.com/drive/folde...UhWMXBLMVl0aUU

[Andrea Alessi]

HI prashantmore,

it's hard to tell without seeing the full dataset, however what you are describing is usually caused by:

- featureless walls, but this doesn't seem to be the case, as you mention that, by themselves, the walls are reconstructed.

- in your case then, i assume there isn't enough overlap between the "connection" of the photos. Try to take more photos, and when dealing with interiors you should move in a parallel way for each wall, and do not stand in one point and rotate. This is the most common acquisition error.

- which zephyr version are you using? with pro/aerial you can fix this also via control points merge

[prashantmore]

I have updated the post and also provided link for dataset. Please let me know if any additional information required.

[prashantmore]

Andrea Alessi :- I am using pro version.

Can you please explain me about control points merge?

[Andrea Alessi]

Simply place control points on an area on an overlapping area between the two walls, reconstruct those walls separately (the two projects must have a certain area in common) and place the same control points in the second project. Merge the project via control points and you're done i suggest you stop for both projects at the sparse point cloud, then do the merge, and then proceed with the dense, this way you can also add an additional bundle adjustment without wasting time.

Make sure to read this https://www.3dflow.net/technology/do...space-merging/ for an example

[Andrea Alessi]

I just saw the dataset, you should probably take few more pictures (some are actually not in focus, i.e. dsc0342.jpg) and avoid the rotation movement only. For example it looks like you moved, stopped and rotated.

Let me explain with the attached image: I did not draw the overlap for clarity reason, but obviously the cameras must have overlap

Also, try to take pictures at different heights, that should help too.

[prashantmore]

Hi Andrea Alessi ,

I tried with control points but I didn't got the output properly.
1st model merged with 2nd model. 1st model is much bigger as compared to the second one.
Do you have any solution for this?

[Andrea Alessi]

Hi Prashant,

it's hard to tell without having a look at the zep file, however from what you're saying it looks like you simply did a workspace merge, not with control points. Try to pick the same control points on both projects before merging them and use "by control points", not "direct merge".

[Scott Nebeker]

Hmmm... VERY few cameras (photos). That's definitely a part of the problem. Keep in mind that anything relying on computer-vision or computer-learning needs LOTS of information. Photogrammetry especially needs overlapping images. There's also that nasty out of focus shot.

Be sure to download the manual for Zephyr. Capturing the best dataset possible is crucial to getting a good model. The part about capturing the appropriate photos and how to do it starts on page 6.

[Scott Nebeker]

I ran the dataset you gave in some spare time and got good results with automatic settings. No control points.

When I gave it control points, the accuracy was REALLY good.

As for the black holes you mention, those are areas that the camera doesn't see or that Zephyr cannot discern with good accuracy. You can choose to have those holes filled by Zephyr.

Notice that the model I created with your dataset is slanted. That's just me making a screenshot of it the way the automatic settings kicked it out. It's easy to correct that orientation but I felt it best to not correct it in order to show that this really is how Zephyr did it.