Request For Feedback: Would you want automatic graceful failover to ORB?

Feedback Ideas and Proposals
1

ORB has been steadily improving in performance and reliability within WebODM as the processing pipeline improves in all areas.

Sometimes we see early OOM (Out-Of-Memory) on memory-constrained systems, or with very large datasets/high quality parameters during Feature Detection and Extraction.

ORB’s binary features are incredibly small in-memory and still very effective in nearly every dataset tested.

So, would you want the processing pipeline to gracefully failover to ORB from any other Feature Type when OOM in early stages to try and ensure you at least get a successful run? (Note: this is similar to how we handle graceful failover from GPU to CPU features, and untiled PointCloud generation to tiled PointCloud generation)

  • No
  • Yes (Always)
  • Yes (Only when --feature-type is default/unset)
  • Yes (Other, in comments)
0 voters
2

Interesting, I’ve never had a dataset OOM during feature detection.

Does this happen with feature-quality ultra?

If that’s the case, perhaps that’s the issue.

1 Like
3

I’m mostly seeing it with MacOS folks who are too memory-constrained, and the pipeline still may OOM later during ortho generation or elsewise, even if we dodge it in the early stages.