← Back to all work Kaggle Research notebook
Solar Filament Grouped CV and Strategy Writeup Public notebook documenting grouped cross-validation and a strategy writeup for solar filament segmentation.
THE PROBLEM
What it sets out Repeated expert annotations of the same solar observation can leak between validation folds, while thin filaments require careful resolution and instance evaluation.
THE APPROACH
How it comes The public strategy writeup groups observations before validation, describes contrast normalization and U-Net experiments, and evaluates instance masks with grouped out-of-fold diagnostics. This is a summary of the author’s documented research workflow.
UNDER THE HOOD
Follow the workflow. Trace the components.
6 COMPONENTS
Solar Filament Grouped CV and Strategy Writeup
Python U-Net pycocotools
Solar Filament Grouped CV and Strategy Writeup / workflow
Solar observations → Grouped validation Solar observations → Contrast / resolution Grouped validation → U-Net experiments Contrast / resolution → U-Net experiments U-Net experiments → Instance masks / RLE Instance masks / RLE → OOF diagnostics Solar observations → Grouped validation Solar observations → Contrast / resolution Grouped validation → U-Net experiments Contrast / resolution → U-Net experiments U-Net experiments → Instance masks / RLE Instance masks / RLE → OOF diagnostics 01 Layer 1 Solar observations Explore ↗
02 Layer 2 Grouped validation Explore ↗
03 Layer 2 Contrast / resolution Explore ↗
04 Layer 3 U-Net experiments Explore ↗
05 Layer 4 Instance masks / RLE Explore ↗
06 Layer 4 OOF diagnostics Explore ↗
Move to tilt · Drag to rotate · Arrow keys Drag to rotate · Tap a component 6 components · 6 connections
Ⅱ Pause motion↺ Reset view
▤ View in 2D2D diagram ◇ Return to 3D
01 Solar observations Explore component + 02 Grouped validation Explore component + 03 Contrast / resolution Explore component + 04 U-Net experiments Explore component + 05 Instance masks / RLE Explore component + 06 OOF diagnostics Explore component + Trace connections 6 relationships Solar observations → Grouped validation Solar observations → Contrast / resolution Grouped validation → U-Net experiments Contrast / resolution → U-Net experiments U-Net experiments → Instance masks / RLE Instance masks / RLE → OOF diagnostics + Select a component to see how it fits into the system.
CAPABILITIES
What’s inside. 01 Grouped validation by physical observation filename
02 CLAHE / disk masking and resolution experiments
03 U-Net ensemble and annotator-consensus strategy
04 Instance separation, RLE and OOF diagnostics
TECHNOLOGY & BUILDING BLOCKS
Python U-Net pycocotools
SCOPE & CURRENT STATE
Public notebook
The inspected notebook is a markdown strategy writeup, not an executable training pipeline.
Experiment results and metric observations are author-reported; no training run or leaderboard result was independently reproduced.
The described native-resolution patch pipeline was marked in progress in the writeup.