How waldstrom/collembot works
COLLEMBOT is a research toolkit for preparing and training Collembola segmentation models, benchmarking alternative backbones, and counting organisms in soil-slide images. The main user workflow runs configured slide images through tiling and filtering, model inference, polygon assembly and fusion, then produces counts, metrics, and visualizations. Training and comparative benchmarking are separate workflows; image upscaling and an interactive Gradio demo are optional capabilities. The inference excerpts are partial, so documented stages are represented, but uncertain internal call links are omitted or marked unverified.
Main entry points: compare_models.py Benchmark Runner · train_model.py Dataset Preparation · run_inference.py Tile & Filter · upscale_images.py Image Upscaling
Components 20
Slide Inference
- Inference Settings
configs/inference.yaml - Tile & Filter
scripts/run_inference.py - YOLO Inference
scripts/run_inference.py - Polygon Assembly
scripts/run_inference.py - Polygon Fusion
scripts/run_inference.py - Metrics & Viz
scripts/run_inference.py
Dataset Training
- Training Settings
configs/train.yaml - LabelMe Data
data - Dataset Preparation
scripts/train_model.py - Data Splits
scripts/train_model.py - YOLO Training
scripts/train_model.py - Model Evaluation
scripts/train_model.py
Model Benchmarking
- Sweep Settings
configs/train_multirun.yaml - Benchmark Runner
scripts/compare_models.py - VRAM Batch Sizing
scripts/detectron_batching.py
Optional Tools
- Image Upscaling
scripts/upscale_images.py - Gradio Demo
scripts/gradio_app.py
Other components
- Researcher
- Slide Images
data/inference-examples - Segmentation Backbones
Connections 21
- Researcher → Tile & Filter: runs inference
README.md - Slide Images → Tile & Filter: provides images
README.md - Inference Settings → Tile & Filter: configures
README.md - Tile & Filter → YOLO Inference: selects tiles
scripts/run_inference.py - YOLO Inference → Polygon Assembly: produces detections
scripts/run_inference.py - Polygon Assembly → Polygon Fusion: assembles polygons
scripts/run_inference.py - Polygon Fusion → Metrics & Viz: feeds results
scripts/run_inference.py - LabelMe Data → Dataset Preparation: provides annotations
README.md - Training Settings → Dataset Preparation: configures
README.md - Researcher → Dataset Preparation: starts training
README.md - Dataset Preparation → Data Splits: prepares samples
scripts/train_model.py - Data Splits → YOLO Training: supplies splits
scripts/train_model.py - YOLO Training → Model Evaluation: evaluates checkpoint
scripts/train_model.py - Researcher → Benchmark Runner: starts comparison
README.md - Sweep Settings → Benchmark Runner: configures sweeps
README.md - Benchmark Runner → Dataset Preparation: reuses preparation
scripts/compare_models.py - Benchmark Runner → VRAM Batch Sizing: uses batch sizing
scripts/compare_models.py - Benchmark Runner → Segmentation Backbones: benchmarks
README.md - Researcher → Image Upscaling: optionally upscales
README.md - Image Upscaling → Slide Images: writes enlarged images
- Researcher → Gradio Demo: opens demo
README.md