How dirac-institute/asteroid_detection_cnn works
ADCNN detects faint asteroid trails by creating labeled image data, training a tiled-image segmentation model, and evaluating predictions at pixel and object levels. Its documented capabilities also include Rubin Butler-compatible data processing and post-processing detections. The graph shows supported calls and documented workflow links; some subsystem implementations were not sampled, so their internal connections are left unspecified.
Main entry points: main.py Training workflow · train.py Training engine · dataset_creation Dataset generation · threshold_scan.py Threshold analysis
Components 20
Data preparation
- Dataset generation
ADCNN/data/dataset_creation - Trail injection
ADCNN/data/dataset_creation/simulate_inject.py - Deterministic injection
ADCNN/data/dataset_creation/simulate_inject_fill_deterministic.py - Rubin processing
ADCNN/data/dataset_creation/pipetasks.py - HDF5 tiled data
ADCNN/data/datasets.py
Model and training
- Run configuration
ADCNN/core/config.py - Segmentation network
ADCNN/core/model.py - Training losses
ADCNN/core/losses.py - Training engine
ADCNN/train.py - EMA weights
ADCNN/training/ema.py - Distributed runtime
ADCNN/utils/dist_utils.py - Training workflow
ADCNN/main.py
Inference and detection
- Prediction
ADCNN/inference/predict.py - Prediction postprocess
ADCNN/inference/postprocess.py
Evaluation
- Trail geometry
ADCNN/evaluation/geometry.py - Object detection scoring
ADCNN/evaluation/detection.py - Pixel metrics
ADCNN/evaluation/metrics.py - Threshold analysis
ADCNN/evaluation/threshold_scan.py
Other components
- Researcher
- Rubin Butler
Connections 21
- Researcher → Training workflow: configures run
README.md - Researcher → Dataset generation: creates data
README.md - Rubin Butler → Rubin processing: provides products
ADCNN/data/dataset_creation/pipetasks.py - Dataset generation → HDF5 tiled data: writes labeled data
README.md - Trail injection → Dataset generation: uses shared helpers
ADCNN/data/dataset_creation/simulate_inject.py - Deterministic injection → Dataset generation: uses shared helpers
ADCNN/data/dataset_creation/simulate_inject_fill_deterministic.py - Training workflow → Run configuration: uses settings
ADCNN/main.py - Training workflow → HDF5 tiled data: loads tiles
ADCNN/main.py - Training workflow → Segmentation network: constructs model
ADCNN/main.py - Training workflow → Training engine: starts training
ADCNN/main.py - Training engine → EMA weights: tracks weights
ADCNN/train.py - Training engine → Pixel metrics: validates model
ADCNN/train.py - Training engine → Distributed runtime: coordinates workers
ADCNN/train.py - Prediction → Segmentation network: runs model
ADCNN/inference/predict.py - Prediction postprocess → Trail geometry: uses component geometry
ADCNN/inference/postprocess.py - Object detection scoring → Trail geometry: labels components
ADCNN/evaluation/detection.py - Object detection scoring → Pixel metrics: reports metrics
- Threshold analysis → Pixel metrics: scans thresholds
ADCNN/evaluation/threshold_scan.py - Threshold analysis → Trail geometry: uses geometry
ADCNN/evaluation/threshold_scan.py - Prediction → Prediction postprocess: postprocesses predictions
- Researcher → Threshold analysis: analyzes detections
README.md