search

seung-lab / neuroglancer

WebGL-based viewer for volumetric data
https://neuromancer-seung-import.appspot.com/
Apache License 2.0
22 stars 10 forks source link

YACN to do list #120

Closed tartavull closed 6 years ago

tartavull commented 7 years ago

Easy stuff

Implement a new data augmentation method, e.g. black out a random block Auxiliary inputs/outputs for floodfilling semantics as inputs affinity map as output Check if regularization helps Play with error localization windowing Support 64-bit segmentation everywhere, and remove assumptions of consecutive seg-ids streamline visualization process (pre-mesh volumes, activate flood-filling on client side, graph server)

Infrastructure

Reduce number of windows in error detection to speed up development Make everything easy to use! document everything! Queue examples on gpu Increase frequency of hard examples Train from sparse/focussed annotations. Switch to precomputed data sources? Set up a grid search for parameters Improve diagnostics for go/no-go decisions for error correction inference

Error Detection

Distinguish merge and split errors More fake errors. In particular, introduce image defects and propagate errors through to segmentation errors Multiscale error detection Training error detection seems to take unnecessarily long. Maybe we should pre-train on reconstruction loss and then fine-tuning on error detection loss. Replace pixel-wise equality with earth mover's distance or Hausdorff distance

Error Correction

Accept flood-filling result only if discriminator approves Choose flood-filling threshold to optimize discriminator approval Generative adversarial training Reduce coverage of error correction Where possible, speed up CPU operations

Regiongraph

Produce a regiongraph with weighted edges instead of binary. This will be essential for Ran's cell splitting

tartavull commented 7 years ago

image

Tensoboard output can now be visualize with neuroboard.py

tartavull commented 7 years ago

train error detection to distinguish different type of errors train the networks to actually distinguish the three cases: obj is a superset of something in human_labels, obj is a subset of something in human_labels obj is empty 1) and 2) could happen simultaneously; this means a merge and split error in the same window.

xiuliren commented 7 years ago

@tartavull yacn do not have blending of output patches?

nkemnitz commented 6 years ago

Doesn't belong here and @jabae said it's not relevant to him, either.