How yu-deeplearning/sayzek-tez works
This repository is a research codebase for aerial/satellite image change detection, centered on PMT-CD: a four-phase progression from binary change localization through transfer/generalization and semantic change to damage assessment. Its principal workflow prepares paired imagery, trains the shared-encoder multi-task model, then evaluates task-specific outputs. Separate branches provide a Siamese-CNN baseline, CORAL domain-adaptation experiments, and reviewer/generalist experiments. The graph follows the README and indexed references; several unsampled subsystem links are omitted where call ownership could not be confirmed.
Main entry points: train_phase1.py Binary training · eval_phase3.py Semantic evaluation · xbd_dataset.py xBD loader · levir_dataset.py LEVIR loader
Components 23
Data handling
- Preprocessing
src/data/preprocessing.py - Image transforms
src/data/transforms.py - LEVIR loader
experiments/pmtcd/datasets/levir_dataset.py - SECOND loader
experiments/pmtcd/datasets/second_dataset.py - xBD loader
experiments/pmtcd/datasets/xbd_dataset.py - xBD mask prep
experiments/pmtcd/datasets/xbd_preprocess.py
PMT-CD model
- Siamese encoder
experiments/pmtcd/models/encoder.py - Feature fusion
experiments/pmtcd/models/fusion.py - Feature decoder
experiments/pmtcd/models/decoder.py - Task heads
experiments/pmtcd/models/heads.py - PMT-CD model
experiments/pmtcd/models/pmtcd.py
Training workflow
- Binary training
experiments/pmtcd/train_phase1.py - WHU pretraining
experiments/pmtcd/train_whu.py - Replay training
experiments/pmtcd/train_generalist_replay.py - Siamese baseline
experiments/baseline/train_simple.py - CORAL adaptation
experiments/coral_adaptation/train_coral.py
Evaluation and experiments
- Semantic evaluation
experiments/pmtcd/eval_phase3.py - Damage evaluation
experiments/pmtcd/eval_phase4.py - Task losses
experiments/pmtcd/utils/losses.py - Task metrics
experiments/pmtcd/utils/metrics.py - Semantic metrics
experiments/pmtcd/utils/scd_metrics.py
Other components
- Researcher
- Aerial datasets
Connections 26
- Researcher → Binary training: starts training
README.md - Researcher → WHU pretraining: starts pretraining
README.md - Researcher → Semantic evaluation: runs evaluation
README.md - Aerial datasets → LEVIR loader: provides pairs
README.md - Aerial datasets → SECOND loader: provides labels
README.md - Aerial datasets → xBD loader: provides scenes
README.md - xBD loader → xBD mask prep: renders masks
experiments/pmtcd/datasets/xbd_dataset.py - Preprocessing → Image transforms: supports
- LEVIR loader → Binary training: feeds training
- SECOND loader → Replay training: feeds replay
experiments/pmtcd/train_generalist_replay.py - xBD loader → Replay training: feeds replay
experiments/pmtcd/train_generalist_replay.py - Siamese encoder → Feature fusion: produces features
README.md - Feature fusion → Feature decoder: combines features
README.md - Feature decoder → Task heads: supplies maps
README.md - Binary training → PMT-CD model: trains model
- WHU pretraining → Siamese encoder: pretrains
README.md - Replay training → PMT-CD model: trains replay
experiments/pmtcd/train_generalist_replay.py - Binary training → LEVIR loader: uses dataset
README.md - Siamese baseline → Aerial datasets: trains on LEVIR
README.md - CORAL adaptation → LEVIR loader: adapts from
experiments/coral_adaptation/train_coral.py - CORAL adaptation → PMT-CD model: adapts encoder
- Semantic evaluation → SECOND loader: evaluates on
experiments/pmtcd/eval_phase3.py - Damage evaluation → xBD loader: evaluates on
README.md - Semantic evaluation → Semantic metrics: computes metrics
experiments/pmtcd/eval_phase3.py - Semantic evaluation → Task metrics: computes metrics
experiments/pmtcd/eval_phase3.py - Binary training → Task losses: uses task losses