/sartorius_cell_instance_segmentation_kaggle

Solution for Sartorius Cell Instance Segmentation Kaggle

Primary LanguagePython

A solution to the Sartorius Cell Instance Segmentation Kaggle

https://www.kaggle.com/c/sartorius-cell-instance-segmentation

Solution summary

  1. Semantic segmentation by Unet.
  2. Instance segmentation by further processing of semantic segmentation with Deep Watershed Transform.

Deep Watershed Transform Network:

Semantic segmentation (Unet)

  1. To generate training target:
python seggit/data/scripts/make_semseg_target.py
  1. To make a training run:
python seggit/training/run_segmentation.py
  1. To make inference:
from seggit.cell_semantic_segmentation import SemanticSegmenter
segmenter = SemanticSegmenter(checkpoint_path='best.pth')
img, semseg = segmenter.predict('sample.png')

Direction Net (DN)

  1. To generate training target:
python seggit/data/scripts/make_uvec.py
  1. To make a training run:
python training/run_direction.py

Watershed Transform Net (WTN)

  1. To generate training target:
python seggit/data/scripts/make_wngy.py
  1. To make a training run:
python training/run_energy.py

Deep Watershed Transform end-to-end (DN + WTN = WN)

  1. To make a training run:
python training/run_watershed.py
  1. To make an inference:
from seggit.deep_watershed_transform import DeepWatershedTransform

dwt = DeepWatershedTransform(checkpoint_path='best.pth')
wngy = dwt.predict(img, semg)

Cell instance segmentation (Unet + WN)

To make an inference :

from seggit.cell_instance_segmentation import CellSegmenter

parser = argparse.ArgumentParser()
CellSegmenter.add_argparse_args(parser)
args = parser.parse_args()
args.pth_unet = 'best_unet.pth'
args.pth_wn = 'best_wn.pth'

segmenter = CellSegmenter(args)

img, instg = segmenter.predict('sample.png')

References