Good evening,
We are using drone images to create 3D of streets. Sometimes the process is failing on the alignment step. We have learnt that one of a possible reason is a wrong flight of a drone. We are trying to define some rules, how the drone should fly. We are planning to use these rules to filter out the images, which are causing the process to fail.
For example, we have learnt, that if the flight has a circula shape, then all cameras, that lie behind the circle might cause the process to fail.
Could you please share with other some other rules, which we could apply to automise the process?
I am looking forward to your reply.
Best regards,
Hello, first, I would like to know what you mean by “failing on the alignment step.” Are you seeing any errors? Does RealityCapture crash? Or do you simply mean that you get multiple components, and all photos are not aligned in a single component?
To get complete components with all cameras, you need to follow the same rules that apply if you are scanning a smaller object with a regular camera on the ground.
To summarize it you need:
- Good Image quality. Images must be sharp, in focus, with deep depth of field, without blur and noise
- Subject coverage. For RC to reconstruct the subject, the entire subject must be captured within the set of images. If there is a gap or missing information in the images, RC cannot fill that gap. For this reason, many images are taken of the subject from differing angles to ensure 100% subject coverage within the images.
- Information overlap. Information overlap is necessary because every single point on the subject needs to be visible in at least two photos in order to be reconstructed. If it isn’t there is no way of triangulating the point. To provide RC with enough information to complete this task, scan images need to have approximately 70% overlap.
We don’t have a dedicated drone scanning material, but we will add it soon. In the meantime, check out the following materials to learn more about the scanning process. As I wrote, the same principles apply when scanning with a camera from the ground or from a drone. The difference is that a drone might be programmed to fly along a pre-defined flight path.
Thank you for your reply, @JakubVanko !
So what we see is the following: the command terminates, but alignment does not reach 100% and when we are trying to export the “.rcalign”, we do not receive the file.
If we remove some of the images and rerun this step, we receive the aligned model. We are trying to find the way, how to filter out these images automatically. The image, attached before, is an example, when it happened. A point is a camera position. It is not perfectly seen, but you could find some images in the red circle, which we had to filter out.
I am looking forward to your reply!
Best regards,
Just for clarification, what do you mean by: “alignment does not reach 100% and when we are trying to export the “.rcalign”, we do not receive the file”? That the alignment is not finished or that not all of the images are aligned?
What is happening, when you are trying to export .rcalign file? How are you exporting that?
I suppose you asked something similar before (to select some images and not use them). Unfortunately, this is not doable in RealityCapture, it uses all imported images.
Looking at your picture, it seems that all images were aligned into one component (as all of them are green).
Dear @OndrejTrhan,
The alignment is not finished. To expor the .rcalign file, we are running -exportLatestComponents command.
When it comes to selecting the images to be used by RealityCapture, we are planning to do the filtering before importing the images. That is why, I believe that previous question is currently out of scope 
I need to say some words to the picture, I have attached. This picture was generated by us using initial metadata of images. The default color is green. There is no additional information encoded by color. The interesting zone of the image is emphasised by red circle. If you look carefully, you will see the green dots there. These images caused the alignment process to fail. I introduce it as an example, of the pattern, we noticed, that the RealityCapture cannot grasp. We hoped that maybe there are other known patterns
I hope that misunderstanding is solved now.
I am looking forward to your reply.
Best regards,
It is really strange that the alignment is not finished. You can try to use the -waitCompleted.) command.
OK, thank you for the explanation., The image looks like the one from RealityCapture’s map view, so I thought it is it.
The main problem there is, that it is not an ideal capture path for photogrammetry, therefore there could be such issues as you described. The ideal workflows are mentioned in Jakub’s post.