Turntable Object Process Flow

[colcrabs]

A question/asking for help with my workflow on building models from a turntable.

My current process for building the full model is the following:

1. Import Photos - 2 Chunks Top and Bottom

2. Batch Process both chunks:

Detect Markers > Align Photos (medium quality) > Build Mesh (medium)

3. Manually remove background and add scale bars.

4. Import Masks from Model

5. Merge Chunks then reprocess:

Align Photos (highest quality) > Build Mesh (highest) > Build Texture

This process has been pretty good and fast and getting me the results I need but I'm positive there are better ways of doing this so I'm happy for any suggestions for ways of improving!  All the objects are pretty small usually no bigger than 10 cm or so.

I'd also be interested in trying to use Mask from Background but I can't seem to get it to work.  No matter what file type, name, type of photo etc. it is it will always come back with the same error message "Can't open file: No such file or directory (2):".  I'm guessing that this would make the process a lot easier.  Another question as part of this, I'm using a light/camera box with the turntable inside.  Would it be possible to build a model of the box and use that as a background mask for each object?  Sorry if that's a dumb question! 

Thanks for any help!

[johnnokomis]

I have a similar workflow but a few changes

1. Import Photos - 2 Chunks Top and Bottom

2. Batch Process both chunks:

Detect Markers > Align Photos (high quality) > remove non-subject tie points > optimize cameras > Build Mesh (medium/high)

3. Remove background and turntable > Build Texture

4. Generate masks from Model

5. Align and Merge Chunks > Optimize cameras again

6.  Build Mesh (high/highest) > Build Texture

I'm not saying this workflow is perfect, it's just what I do. Open to any suggestions.

[colcrabs]

Thanks for responding!

When you remove non-subject tie points is that manually removing them or is there a process for that?

Also, for optimizing cameras, do you calibrate your cameras first?  I think I have really rough and ready processes that should be refined a bit.

[renevg]

Hi colcrabs,
I pretty much have your workflow and it's working for me.
I try to get the best input possible; this is key.
estimate image quality to keep the best.
detect markers at each chunk and check if markers are ok in photos.
align photos low or medium quality each chunk (batch)
build model and use bounding box to cut away pieces not needed (lost part of object will be in other chunk)
create model and generate masks from model
If you have photoshop you can skip modelmaking by building your masks there (both shape of model and focus blur can be masked there and merged with existing masks in metashape)
check masks
merge chunks. Markers and scale bars will be in new chunk.
start over with your photos with the masks to keypoints: align, high quality
Optimize your tie points:
1.   Reconstruction uncertainty
2.   Projection accuracy
3.   Reprojection error
mind you: deleting tie points any different way than through optimization doesn't work. They will be back in dense cloud.
build your product; generate depth maps, dense cloud,
you can clean up dense point cloud by filtering: filter points by confidence I usually use.
generate mesh
check your mesh with model>confidence. preferably darkblue. If so: be happy. Red or yellow parts? Consider taking new photographs and add them to the project.
build texture if needed
Every project has it's own way of best workflow in my opinion.
good luck.

[johnnokomis]

For tie points I'll first click on Model>Gradual Selection then use those options to select the poor ones. Then remove the others manually. This is just so when you optimize the second time, on the merged chunk, it only uses those for optimization.

One step I forgot is to remove the markers from each chunk before aligning them. Not sure if that makes a difference but I do it since the same marker numbers are in both chunks.

[renevg]

I'd like to add to my comment in reaction to johnnokomis:
keeping markers is essential in this workflow. This gives scale to the project. Make sure you have markers which are only numbered once in the first chunks so there is no overlapping. With turntable I use coded targets that get detected with 'Detect markers' and get a number. (don't use same targets for both chunks). You can print metashape targets or buy commercial ones. check the known distance between the targets and use the distance for your scale bar. That way you've got images with just the object + targets/markers with known distance in your merged chunk. 

[jrp]

I do some very similar work to what you are doing, and I use a variety of workflows similar to those discussed on dificult objects that won't behave, but I actually find that the following works more often than not with good photo sets:

1) import photos all into a single chunk
2) align
3) build dense cloud
4) manually chop out the unhelpful bits.
5) process as normal with any degree of optimisation you feel liek that day

Step 4 benifits from propping the object up off the turntable factionally with little bits of foam etc, well underneeth the base of the object so that camera can't see them.

When the object isn't quite right, or there is minor misalignment, I sometimes find that this model is close enough for masks from object before starting again.

If I remember right, generate masks from background is designed for dealing with situations (for example) where you take a photo of the turntable without the object on it, and use that as the input. It takes the colour diference between "background image", and the actual image, masking out those that are similar. With a turntable, it will usually get a little confused that the turntable is moving, but it will do a good job on the wall behind and an ok job on the turntable. good oclour contract between object and background is needed otherwise you risk bits of the object vanaishing.

I believe it's asking for a folder containing the background image file, then separatly a filename within that folder -- this is so that other workflows where the background iamge is diferent for every image can be managed.

A while back we used to use this to create a primative "background from how close to white the pixel is" by feeding in a pure white image of the same resolution as the images, and tweaking the tolerance.

[Alexey Pasumansky]

If you are still getting problems with the background masking, please share the Import Masks dialog screenshot with the filename template that you are using and specify the names of the original images and background image files.

As for the actual workflow question, then depending on the object type you may be able to process all the images at once in the single chunk, without a need to generate the masks for the subsets of images. Such approach has been covered in the following tutorial article:
https://agisoft.freshdesk.com/support/solutions/articles/31000163268-turntable-capturing-scenario-and-processing-workflow