Multiplicated shifted geometry

[TheoNanu]

Hi,

First of all, congratulations on this work, and thank you for sharing it with other computer vision enthusiasts and practitioners.

I have conducted multiple experiments using this model to understand its capabilities for precise measurement of different objects’ geometric features.

I have a couple of questions about the results I have obtained:

From my experiments, it seems that using multiple views results in a shifted, duplicated geometry of the same object, even when high-confidence filtering is applied (see the point cloud image below; what you see there are two different electrical pins of a connector, and the pin tip appears doubled, with the second instance slightly shifted from the first).

For the same connector, when only one view is used, this effect does not occur:

I also tried generating with two images from the same view but using different exposure settings, and this layering effect becomes even more obvious (the first image shows a planar surface that appears doubled, and the second image shows an electrical pin):

The question is, why is this happening, and what can be done to prevent it?

One more question I have is whether the scaling applied to the dimensions is the same on all three axes, or if the scale differs along each axis?

Also, do you have any advice on how to get the most out of this model for tasks such as measuring the geometry of objects?

Thank you in advance for your help.

[jytime]

Hi currently I think this is the known limitation of the feedforward models, i.e., the alignment may have small mismatches. Did you try the reconstruction with the BA option enabled? With post optimization like BA, this problem should be mitigated, although BA takes longer time.