GPU and decreased geolocation accuracy

Hi everybody, I would like to share some findings of benchmarking the use of GPU vs no GPU. We were using VMs (Google Cloud Engine) with 16 cores and 80 RAM for that task (survey of 0.25ha with 9.5mm GSD). Here a resume of our findings together with some graphics:

  • Orthomosaic creation with 1xGPU was 65% quicker (1:42 vs 2:38 hours).
  • Orthomosaic creation with 1xGPU was 250% more expensive (US$ 0.75 vs1.87).
  • The use of 1xGPU decreased Relative Horizontal Accuracy CE 90 in 51% (3.2 vs 6.3cm).
  • The use of 1xGPU decreased the # of Reconstructed Features in 23% (15.703 vs 12.073).

    I would like to get some feedback from the power users, if they experience similar behavior when using GPU(s). Is there a specific ODM setup to follow in order to avoid a decreased # of keypoint matching with GPU?
    Feedback is greatly apreciated!
    Best,
    Gabor
1 Like

Possibly related: SIFT_GPU ignores min-num-features ? · Issue #1437 · OpenDroneMap/ODM · GitHub

2 Likes

Hi Piero, thank you for your quick answer. We were using feature-quality: medium. Changing the option to high would still produce a decreased # of keypoint matching when using GPU compared to no GPU?
Best,
Gabor

1 Like

Given what we’ve observed, likely yes, but it should also still improve the final product quality.

1 Like

Thanks for sharing, Saijn!

Best,
Gabor

1 Like

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.