TL;DR:
Metashape processing shows significant variance differences between Windows and Linux for alignments of frames from videos taken with f2.2 and f1.8 lenses. Linux has minimal variance, while Windows shows high variance in camera errors and f-values, especially for f2.2 videos. Why? How to fix?
I am trying to process smartphone videos taken with different lenses with the Metashape Python API (2.1.2), and I noticed significant differences between Windows and Linux. The processing machines are identical apart from the processor, which is one generation newer on the Linux machine (see details below). One batch of videos (Video 0,1,2,3) is taken with an f1.8 lens at 1920x1080 pixels resolution, and the other batch (Videos 5,6,7) is taken with an f2.2 wideangle lense at 4000x3000 pixels resolution. I extract a frame every 0.2 seconds and assign the RTK-enabled geolocation to each frame, then I run the alignment of the photos in Metashape, using the parameters below for each run. I repeat the experiment (addPhotos, matchPhotos, alignCameras) 15 times for each video, so 15 runs per video on Windows, 15 runs per video on Linux. I am investigating the variance inbetween runs regarding camera errors and estimated internal camera parameters. I am letting Metashape estimate the internal camera paramers freely, providing no reference information. The results are intriguing. I posted a graph in the attachment. There is a significant difference between Linux and Windows. On Linux, the videos show no variance in camera errors at all, whereas on Windows, the errors for the f2.2 videos are highly variant. On Windows, the f-value estimate for the f2.2 videos is highly variant, while the f-value for the f1.8 videos is barely variant. On Linux, the f-value shows almost no variance. I am not very knowledgeble about the other paramers, but in general they show a higher variance on Windows than on Linux. Can someone elaborate on why these platform differences would occur, and how I could mitigate them?
Alignment Settings:
Accuracy: High
Generic preselection: No
Reference preselection: Source
Key point limit: 80,000
Key point limit per Mpx: 1,000
Tie point limit: 5,000
Exclude stationary tie points: Yes
Guided image matching: No
Adaptive camera model fitting: Yes
System Information:
CPU: Intel(R) Core(TM) i9-14900K (Linux) and Intel(R) Core(TM) i9-13900K (Windows)
Memory: 128GiB
GPU: NVIDIA GeForce RTX 4090
Topic: Alignment differences between Linux and Windows (Read 2597 times)
