segmentation_benchmark

Code for: Content-aware segmentation of objects spanning a large size range: application to plankton images

Thelma Panaïotis

PhD student

Laboratoire d’Océanographie de Villefranche (UMR 7093)

---

Benchmark for three segmentation pipelines on ISIIS data:

• threshold-based pipeline: adaptive gray threshold
• T-CNN pipeline: CNN-based bbox proposal for small objects + thresholding
• T-MSER pipeline: maximally stable extremal regions

Structure

data contains all data for segmentation benchmark:

• manual: ground truth data
  • stacks: manual stacks
  • segmented: manual particles for Ecotaxa import from manual stacks
  • particles: manual segments generated from manual stacks
• regular_apeep: apeep output for threshold-based pipeline
  • segmented: segmented images from apeep threshold-based pipeline
  • particles: particles from apeep threshold-based pipeline
• semantic_apeep: Apeep output for T-CNN pipeline
  • segmented: segmented images from apeep T-CNN pipeline
  • particles: particles from apeep T-CNN pipeline
• mser: output for T-MSER pipeline
  • mser_measurements.csv: properties of T-MSER particles
  • mser_matches.csv: matches of T-MSER particles with ground truth particles
• raw_frames: raw frames from avi files (generated by 00.get_raw_frames.py)
• matches_bbox: particle matches (generated by 03.match_particles.py)

lib contains needed scripts.

Scripts

• 00.get_raw_frames.py: extract raw frames from avi files for benchmark images
• 01.process_manual_stacks.py: process manual stacks by extracting and measuring particles for Ecotaxa import, and generate segmented images
• 02.extract_manual_ecotaxa.R: extract manual particles with taxonomy from Ecotaxa, locate them in avi files
• 03.match_particles.py: match manual particles with those from apeep threshold-based and T-CNN
• 04.matches_stats.Rmd: compute global precision and recall, precision and recall per size class and recall per taxonomic group for all three segmentation pipelines

Results

A benchmark report containing computed statistics is generated: 04.matches_stats.html