inaccurate meshing

[TR7]

Describe the bug
inaccurate/broken meshing. I guess it's not supposed to be like this

To Reproduce

• download dataset (162 Photos): http://hosting141203.a2e6d.netcup.net/Thomas/HDKornmarkt/HeidelbergKornmarkt.zip
• start meshroom 2019.2.0
• load images per drag&drop
• set initialization mode to: unknown (otherwise Structure From Motion fails)
• set SIFT describer preset to: medium
• force cpu: off
• click on "start"
  all other parameters are the defaults

Expected behavior
meshing should be a way better with the default values and so many pictures.

Screenshots

overview camera poses:

Logs
http://hosting141203.a2e6d.netcup.net/Thomas/HDKornmarkt/Logs.zip

Desktop (please complete the following and other pertinent information):

• OS: win 10
• Python version 3.7
• Meshroom version: release 2019.2.0

[fabiencastan]

Look at the tooltip on the orange flag on your images. The camera model is probably missing from the sensor database.

[natowi]

Try:

GoPro;FUSION;6.17;dpreview

If it is this model

Tested on nine images:

[TR7]

i tried this model (GoPro;FUSION;6.17;dpreview) with following results:

i also tried with only 9 images, like you have done, and that works without problems.

it seems the problem occurs with more than one loop around the target object. also it seems like there are 3 layers of ground-floor-point-clouds in the SfM output instead of one.

i know, it will take some time to compute, but it would be a very helpful information for me if the reconstruction with all 162 images causes the same problems for you.

[TR7]

good news, it seems to work with more SIFT-Points AND "min observation for triangulation = 3"

[fabiencastan]

@TR7 Do you have the other side of the gopro fusion 360? If yes, could you share it? It would be interesting to declare it as a rig of 2 cameras and see if it improves the results.

[TR7]

one the other side from the gopro, there is me standing in front of the camera, so not really useful. but i will take some other pictures and can of course share them (both sides with me only on the side).
is there a place or do you have a system, where to collect all the links to test-datasets? or just post it here?

[natowi]

Here I proposed to collect user contributed datasets similar to the Monstree demo dataset to test 360 and fish eye images.
@fabiencastan can we create a new Repository under Alicevision to collect a few small test datasets?

[TR7]

whats your prefered amount of front/back photos per dataset? special kind of objects/places?

[fabiencastan]

@TR7 It would be ideal if you could make an indoor and an outdoor dataset.
Share a link here and we will see if we should store it as a reference dataset for this use case. I'm not sure how to organize that for now.

[TR7]

another outdoor dataset with 2x 128 JPGs from GoPro Fusion (both sides, 0.5 Sec Photo Timelapse):
http://hosting141203.a2e6d.netcup.net/Thomas/Scans/08/rig08_2x128Images_Outdoor_GPFUSION.zip

[TR7]

an here a small indoor dataset for test purposes:
http://hosting141203.a2e6d.netcup.net/Thomas/Scans/06/rig06B_FrontBack_2x21Images_Indoor_GPFUSION.zip

@fabiencastan
just tell me, if you need more or bigger datasets.

[fabiencastan]

@TR7 Thanks a lot! I see that you have already organized them as a rig of 2 cameras. Did you already tried to reconstruct them? And compare with/without using the rig?

[TR7]

@TR7 Thanks a lot! I see that you have already organized them as a rig of 2 cameras. Did you already tried to reconstruct them? And compare with/without using the rig?

yes i tested a lot with over 14 datasets from the GoPro Fusion.
I had several problems. Some of them I could solve or avoid by myself:

1. when importing the EXIF images into Meshroom they must be at least 1 second apart, otherwise they will be skipped. I wrote a Python script which automatically changes the recording time strings. Probably the problem arises because Meshroom creates an image identifier based on the name and time (second-based)?

2. the serial numbers in the EXIF tag must be different (but the GoPro Fusion creates the same serial number in the EXIF tag for the back and front sides). I wrote a Python script that inserts a "B" or "F" before the serial number string in the EXIF tag.

Some other (more about the rig) problems are unsolved yet:

3. camera rig constraint does not prevent the camera pairs from drifting apart [question] #906

4. in some datasets: with 1000+ images the reconstruction only works without "Use Rig Constraint". Otherwise it fails in SfM. But i'm still looking into it, though.

5. so far I haven't found out why sometimes the rig symbol (see marker on the 1st picture) is shown after the import and sometimes not.

[natowi]

@TR7 Can I add a few of your images to https://github.com/natowi/meshroom-360-datasets under CC-BY-SA-4.0 license?

[TR7]

yes!