Significant differences in point-numbers of the same basic material

[FHofer]

Hello,

let me give a short summary of our situation. We are using frames extracted from videos to reconstruct areas that are often poorly lit. Thus, we have come to anticipate the generated results differentiating from “normal” expectations. As a countermeasure, we usually run 4 to 12 generating processes for each set of video frames in order to get an overview on what results are good/bad and what we can expect on average.

This is where my problem comes in. Yesterday, I processed multiple sets of frames (each four times). The spare point cloud numbers of a specific one where 29,165, 30,080, 11,415 and 30,120 points (with the third result clearly being one of the reasons why we process the frame sets multiple times).

Today, I repeated the process another five times, but with no other frame sets being processed. The resulting sparse point cloud numbers were 24,100, 23,763, 18,096, 23,482 and 11,021 points. The same parameters where used, the only difference being that the four repetitions yesterday where processed along with multiple others in a batch process, while the five repetitions today where processed in a batch process of their own.
All parameters of the alignment and generating processes are set to the highest levels, except the accuracy of the “align photos”-process in order to avoid manipulation of the frames through upscaling.

Is there an explanation for this variation of the numbers of the sparse point cloud points?

Thanks in advance,
PO1989

[Alexey Pasumansky]

Hello FHofer,

Camera alignment operation is using stochastic element, so each run you will observe different results. If the surface/object of interest is well textured and the image overlap and quality are good, then the observed alignment difference will be minimal, but for cases of unstable alignment (lack of overlap, insufficient texture features, low image quality) the alignment difference would be considerable.

I can also suggest to check, if the results do provide more or less similar parameters of the camera calibration. Since video frames are lacking the EXIF data, it may be also reasonable to input some approximate focal length values to the Initial tab prior to the alignment run (either single value in pixels, or focal length and sensor pixel size in mm).