Reconstruction with +4 thousand photos from at least three different cameras

[santfloyd]

Hello, I´m carrying out a reconstruction of an entire building using photogrammetry. I have three different sensors and prior to 2025 release, in the 2023 version, I was using augmented reconstruction to handle intrinsics calibration per sensor and also a gradual reconstruction, but at some point that approach couldn´t achieve the whole interior-exterior reconstruction. So which strategy do you recomend:

• Dropping in to a single node all images (interior-exterior) from one camera and then manually creating an augmented reconstruction pipeline with other set of photos (interior-exterior) from another sensor and the doing the same with the last sensor

• Carry out the gradual reconstruction hoping that with this 2025 version the reconstruction will succed

• dropping into a single node all 4 thousand images and the compute

Thanks in advance

[natowi]

What is your capturing process? What kind of sensors, how many, in which configuration?

Is this a structured reconstruction (systematic capturing) or more random image order?

What is the overlap between interior/exterior for example?

I think I would split the dataset, image matching will be improved by that, as it does not make sense to try to match image 20 to image 4003 if it will be in a totally different area.

I would create one group per room / area and have enough overlap. Then I would test different alignment approaches, Meshroom offers multiple nodes that allow to merge/align the data at different stages of the reconstruction.

[santfloyd]

I used two drones and one handheld camera for ground imagery. The configuration is random, since I was just trying to cover as much detail as possible for the small halls and other features. I include the camera positions

The overlaping is approximately 70% on average. By splitting the dataset you mean using an "agmented reconstruction" structure nodes, like in the previous version of Meshroom or to process in independent pipelines and then merging the data, at which stage do you recommend merging the data?

Thank you in advance

[natowi]

This is a rough outline: I would create a pipeline for the exterior reconstruction and one for the interior reconstruction and one that has a good overlap between drone images and handheld images, then use the common pictures to align the interior and exterior reconstructions using the alignment node.

[santfloyd]

I will give it a try to the method you propose, I asume using the SfMalignment node. But I still wonder if I still have to import images fron different sensors in chained "augmented reconstruction" pipelines, one for each sensor, or I can drop in the same camerInit node all images form different sensors and Meshroom 2025, is smart enough to compute the respective intrinsics for each camera. I will update as soon as I get some results after computing. Thanks

[natowi]

You will need to do some tests to find the best setup for your workflow