Use Case:
We want to setup a workflow where predictions on a full volume can be proofread (i.e. locally corrected and/or marked as correct). The training pipeline, in particular the sampling, needs to be adapted for this data pattern.
Current Sampling
Crops define valid sampling regions, RandomProvider randomly chooses a crop (with a probability based on the crop size) and RandomLocation picks uniformly within a crop.
RejectEfficiently exists in corditea but is wasteful (loads data then throws it away)
Goal
Support either:
- definition of proofread regions of interest via metadata
- probability map based sampling over whole volumes for more flexible region selection (ideally with options to create probability maps on the fly)
Key advantage is that samples can extend freely beyond the region since data exists everywhere. Therefore edge artifacts for the label processing can be minimized
Design
New node in corditea: saalfeldlab/corditea#1
Use Case:
We want to setup a workflow where predictions on a full volume can be proofread (i.e. locally corrected and/or marked as correct). The training pipeline, in particular the sampling, needs to be adapted for this data pattern.
Current Sampling
Crops define valid sampling regions,
RandomProviderrandomly chooses a crop (with a probability based on the crop size) andRandomLocationpicks uniformly within a crop.RejectEfficientlyexists incorditeabut is wasteful (loads data then throws it away)Goal
Support either:
Key advantage is that samples can extend freely beyond the region since data exists everywhere. Therefore edge artifacts for the label processing can be minimized
Design
New node in corditea: saalfeldlab/corditea#1