Image Dataset of Full Body 88 images for 3D Scan

[Yukihiro Iwayama]

I’m releasing image dataset of full body scan Photogrammetry.

https://github.com/HappyOcean/Full_Body_Photogrammetry

dataset sample

88 images - 8MP, Raspberry Pi Camera Module v2

One of the reconstructed and textured resulting 3D model is skfb.ly/6NUVq on Sketchfab

Yukihiro_Iwayama_Full_Body_Photogrammetry

Directly uploaded from RealityCapture no modification in other 3D software.

Needless to say, you can use these images Metashape. the initial choice was just due to my convenience at this moment. I use Metashape as well though I'm new to this forum.

Free to download and use. Give credit to : Yukihiro Iwayama : https://twitter.com/HappyOcean, Yukihiro Iwayama : https://www.facebook.com/yukihiro.iwayama

Hope this could contribute as some sort of reference data set especially for those who are not quite familiar but are interested in body scan.

There is too little of this kind of full image date set for A Pose Human. Personally, I haven’s seen it that’s why I release.

[Yukihiro Iwayama]

Retopology by Wrap3
https://skfb.ly/6NWU8

[Mak11]

Thanks for sharing. Unfortunately Metashape is really bad with those kinds of data-sets compared to Reality Capture. There's no way to get anything good out of without  masking every single photo (even though camera alignement is perfect Metashape is unable to detect enough feature on the subjects to properly reconstruct it. 80% of the Tie Points are on the background.





Maybe Alexey can help...

Mak

[Yukihiro Iwayama]

Mak, Thank you for your testing and feedback as I was doing my way and getting nowhere. I'll try masking, I guess that the upper body could be challenging anyway.

[Mak11]

Yukihiro,

Yeah this is literately one of the situation where Reality Capture is simply unbeatable. For whatever reason Metashape (& Photoscan before) simply can't get those type of scans right unless masking is used. Unlike RC which doesn't require anything to get a superbe mesh. 😒🤨

You should email Alexey & the Agisoft team at support@agisoft.com and maybe asked them what would be the best way to get anything out of this data set.

Cheers

Mak

[Alexey Pasumansky]

Thanks for sharing. Unfortunately Metashape is really bad with those kinds of data-sets compared to Reality Capture. There's no way to get anything good out of without  masking every single photo (even though camera alignement is perfect Metashape is unable to detect enough feature on the subjects to properly reconstruct it. 80% of the Tie Points are on the background.

That's what I've got using depth maps based reconstruction (no masks applied).

Although I see some defects, they are not so severe, like on your screenshots.

[Mak11]

Alexey,

Can you share your settings ?

Here are mine (depth maps method with no masks):



And here's with everything on medium:



Cheers

[Mak11]

Alexey,
I've tried again with different settings and still can't get anything close to what you are getting.

Cheers

Mak

[bestable]

Me neither. I would like to see detailed walkthrough.. For now it seems that reality capture performs a lot better.

[Mak11]

Glad to know that I'm not the only one unable to get anything usable.
Would really like to know which settings Alexey used.

Another failed attempt..



Cheers

Mak

[Alexey Pasumansky]

Hi,

The steps that I have followed:
- aligned using High accuracy, no preselection, 40 000 / 10 000 for point limits,
- detected markers,
- optimized the camera alignment,
- build mesh (High quality) for the whole (default) area,
- reduced the bounding box to the area of interest and re-built mesh have re-use depth maps option enabled.
After the reconstruction used Smooth Mesh operation to reduce the slight noise on the surface.

[bestable]

Any workflow for Standard version users?
I feel left out a bit.

[Mak11]

Hi,

The steps that I have followed:
- aligned using High accuracy, no preselection, 40 000 / 10 000 for point limits,
- detected markers,
- optimized the camera alignment,
- build mesh (High quality) for the whole (default) area,
- reduced the bounding box to the area of interest and re-built mesh have re-use depth maps option enabled.
After the reconstruction used Smooth Mesh operation to reduce the slight noise on the surface.

Thanks Alexey

Detected Markers ? Isn't this Pro only feature ?

Cheers

Mak

[bestable]

That is my progress pics of the described workflow except i did not use "detected markers" part:

[Mak11]

Alexey,

Getting better results (the trick was obviously to generate the mesh once for the whole volume to get as much depth maps as possible before resizing it down..& re-using the depth maps..)

As usual the mesh is plagued by the same issues as introduced in 1.5.1 (holes, unwanted polys inside, non connected polys etc)

This is with main/mesh_trimming_radius : 0





main/mesh_trimming_radius : 30





Cheers

Mak