Aligning large dataset

[E_Verweij]

Hello all,

I am trying to create an orthomosaic of two datasets: one containing tiff files with one channel, and one containing jpeg photos of RGB data. I have trouble to correctly align the cameras. The datasets are large, they contain +11,000 images. 

The datasets have been clustered with each cluster containing approx 500 photos, so we can process smaller datasets. Still not all cameras are being used, and the result is distorted. Also there seems to be difference between the quality of the alignment of the tiff-files compared to the jpeg-files.

All this left me with the following questions:
- What is the influence of the hardware/software/data in the alignment process? Does a more powerful computer improve the quality of the result or only speed up the process. Does the shape of the data have an influence?

- Is there a limit of the nr. of photos one should use, and in what way is the aligning process influenced by the number of photos.

- What is the difference in aligning JPEG files vs TIFF files. Result of aligning 500 tiff files, usually gives a good result, but with JPEG files I can only get a reasonable result when using 100 photos.

A better understanding of the influence of these factor would help me a lot, so any help will be appreciated.

[Micha]

Sounds like the quality of your JPGs isn't sufficient. Maybe they are compressed too much / lost too many details. TIFF for example is an uncompressed file format which stores every detail of the image. Can you post some example images of your project?

[E_Verweij]

That might explain the difference between the TIFF and JPEG files.

What is troubling me more is that when I process a cluster of say 100 files, I get a good result, but when I increase the number of cameras to 200, the photos are not correctly aligned.

Also when I run the same dataset multiple times, the differences are big: one time all 100 cameras are being used, when the next time only 35 cameras are aligned.

I need to understand what is causing this ambiguity

[dewang]

Hi,

We are aligning a huge data-set 50000+ images of 20MP each

Computer Specs: i7 8600k 3.6Ghz ; 64Gb Ram; Graphics 2080 11Gb

The alignment process is taking too long for single chunk. It finished Matching photos in 4 hrs and completed 98percent Alignment process in 16 hrs.
But its already 4 days now, still it is processing at 99 percent, time remaining is shown as 30mins but is doesn't complete in that time.

Has anyone else faced this same issue?

[Mak11]

dewang

Unfortunately you still have time to take a 2 way trip to Jupiter & back to earth and it probably won't have finished. 50000+ 20MP photos on this PC config in a single chunk is not going to cut it. Which Quality setting did you use (Low/Medium/High/Ultra ?)

Alexey will probably have better insight though.

Mak

[Alexey Pasumansky]

Hello dewang,

Can you send the part of the latest processing log lines from the currently working node?

Also please specify the version of Metashape Pro used and, if possible, recall the following:
- was the adaptive camera model fitting option enabled in the Align Photos dialog?
- was rolling shutter compensation enabled for the camera calibration group in the Camera Calibration dialog?
- which accuracy settings for cameras locations and orientation angles have been used in the Reference pane (default or loaded from the image meta data)?

[dewang]

Hi,

Thank you for your reply.

Please find attached the latest processing log lines of the currently working node.

Version of Metashape - 1.5.1 (64 bit)

As I can recall following were the options:

Adaptive Camera Model fitting was OFF
rolling shutter compensation was disabled
Accuracy setting for alignment was set to High, with generic and reference check box ticked.
Match Points 40000 and Tie Points 4000
Accuracy settings for camera locations and orientation angles were default

Total 76441 images to be precise (mistyped in earlier post)

Also attached is screenshot of current processing.

The last error was of Bad Allocation which was due to RAM of another Node (32GB)

[Alexey Pasumansky]

Hello dewang,

Can you also provide the actual log from the current node? (via Agisoft Network Monitor right-click on the node and select "Details" option).
It is it the insufficient memory issue, it could be an infinite loop for the final step when the same sub-task is being processed for the long time until failing due to memory issue.

But if you see frequent "merging failed" lines in the log, then I could suggest to pause the Batch and without re-starting the server just update Metashape version to 1.5.3 on the nodes and re-connect them and try to finishe the job (run 64 GB node at first).

[dewang]

Hi,

Thank you for your response.

Please find attached actual log from the current node.

Okay, will trying the version

[Alexey Pasumansky]

Thanks for providing the log.

I hope that updating the version of Metashape Pro to 1.5.3.8469 should allow to complete this task.

You do not need to update the server version, just put the task on pause, disconnect all nodes and run the latest version on the nodes.


If the processing finishes on the version 1.5.3, I could suggest to re-align the set using the available matching points in the same version again and compare the results with 1.5.1. As we expect the results acquired in the latest version should be more accurate.

[dewang]

Hi,

That did not solve the issue, getting the same merging failed error.

Attached is the recent log for reference.

[Alexey Pasumansky]

Hello dewang,

Are all the nodes now running the version 1.5.3.8469?

I could suggest then to cancel the task and run align selected cameras operation in the latest version, as the tie points should be kept in the project and you do not need to perform the matching again.

[dewang]

Hi,

Yes. All Nodes are on same verion 1.5.3.8469

I did cancel the task, and re-run the process. But same error.

Downgrading will help?

[Alexey Pasumansky]

Hello dewang,

Please send the processing log from the new version nodes to support@agisoft.com (in the same ticket thread that you have already created).

Also please specify, if the project files folder with the tie points only is not big enough and whether it can be transferred to us, so that we could reproduce the problem and check why it is still there?

[dewang]

Hi,

Unfortunately, as the newer version didn't work so we cancelled the processing and we shared the project with one of our friend who has Photoscan 1.4.3.

The Alignment process got completed in 24 hours on following PC (i5 8600k; 64 GB Ram; RTX 2080 and GTX 1050Ti ; on network of 3 PC)

At this current stage, we have completed the Alignment and inserted GCP information.

Now we would like to proceed further on network processing setup of 4 nodes with Photoscan 1.4.3
As we are completely new to Agisoft, we would like to know how to proceed further?
Should we Split in chunks and then proceed with Dense Point Cloud and Mesh? What should be the overlap percentage?

After processing the data in separate chunk, how should we merge them and proceed with Ortho and DEM generation?

It would be really helpful if you can guide us.