Agisoft Metashape
Agisoft Metashape => General => Topic started by: Paulo on April 20, 2021, 01:56:33 AM
-
Hello community,
once again I compared alignment results on same 48 image data set using same parameters and same gradual selection process for v1.7.3 vs v1.6.6 using following system:
OS Windows 64 bit
RAM 15.95 GB
CPU Intel(R) Core(TM) i7-3630QM CPU @ 2.40GHz
GPU(s) GeForce GTX 670MX
Results are shown in following text file. It is clear that total alignment time for v1.7.3 is much larger than for v1.6.6 (7 min 44 s vs 5 min 26 s or 42 % more). The 2 alignments have same accuracy but version 1.7.3 does generate more tie points (121 000 inicial vs 89 000). Unless there is some optimization in processing times for 1.7.3 (including new depth maps generation times), I will continue to use v1.6.6 as productivity is much greater on such hardware systems....
-
Hi Paulo,
1.7.3 has generated 36% more tiepoints than 1.6.6 so it makes sense that 1.6.6 is faster for you.
What you could do to improve your 1.7.3 alignment time is to limit the amount of Key & Tie points to for example 60 000 and 20 000 respectively.
For my workflow, anything above 60 000 Key points and 20 000 Tie points is a waste of processing time for little to no extra details
Regards,
Martin
-
Good point, MartinVw,
however I am just using 30 000 kpt and 10 000 tpt limit. But I wonder why 1.7 finds more tpts while the result accuracy wise is same for 2 versions....
-
Hello Paul,
Can you share the dataset that you are using for testing? Or specify the resolution of images used and describe the project type (aerial, close range).
Please also confirm, that GPU has been utilized properly for 1.7 run (no errors in the log).
-
Hello Alexey,
The project tested consists of a 48 image data set from Phantom 3 camera (12 Mpix) at 50 m height above ground. As suggested by Martin, i redid alignment on 1.7.3 but using tpt limit of 6 000 instead of original 10 000. And I got basically same amount of inicial tpts (87 000) as 1.6.6 test (with 10 000 tpt limit) and also total alignent time of 5 min 46 s vs 5 min 26 s for 1.6.6:
Tie Points
Points 32,106 of 87,545
Point colors 3 bands, uint8
Key points No
Average tie point multiplicity 3.4343
Alignment parameters
Accuracy High
Generic preselection Yes
Reference preselection Source
Key point limit 30,000
Tie point limit 6,000
Exclude stationary tie points No
Guided image matching No
Adaptive camera model fitting Yes
Matching time 1 minutes 38 seconds
Matching memory usage 580.21 MB
Alignment time 4 minutes 8 seconds
Alignment memory usage 52.78 MB
Optimization parameters
Parameters f, b1, b2, cx, cy, k1-k3, p1, p2
Adaptive camera model fitting No
Optimization time 2 seconds
Software version 1.7.3.12136
File size 7.87 MB
No GPU errors appeared during alignmemnt... see attached logs for alignment on 1.7.3 with 30 000 kpt/6 000 tpt limit and same alignment on 1.6.6 with 30 000 kpt/10 000 tpt limit
PS. If you want to test on your side, i sent the link to image set to support
-
Hello Paul,
You may get somewhat different timing, if you disable "use CPU" option in the GPU preferences tab.
For this dataset I've got about 25-30 seconds matching time in 1.7.3 compared to 15-25 seconds in 1.6.6 (have checked several consecutive runs in the same Metashape instance) with RTX 3080 graphic card, using the same key point / tie point limits as in your initial tests. And alignment time 50-60 seconds (in 1.7.3) compared to 40-50 seconds (in 1.6.6) on quite old desktop with i7-4960X CPU.
-
Thanks Alexey,
so you are still getting relatively higher alignment times for 1.7.3 vs 1.6.6. Is that normal? And if algorythm has not changed, why does 1.7.3 find more initial tpts than 1.6.6?
In my experience, using low performance graphics card as GTX 670 MX, processing times are lower when enabling CPU while GPU processing. Is that your experience on your side?
Again for your insight on this issue...
-
Hello Paul,
In 1.7 there were changes both in image matching and alignment that should improve the stability of alignment, especially in previously unstable cases.
-
Ok,
then I can understand the different times and quantity of Tpts determined using same parameters for both versions...Thank you for the clarification!