How cubicasa/cubicasa5k works
CubiCasa5K is a floorplan-analysis dataset and multi-task neural-network project. Its user workflow turns annotated SVG plans into labels (optionally caching samples in LMDB), loads and augments image/label samples, and trains or evaluates a model that predicts room/icon segmentations and heatmaps. Evaluation includes metrics and polygon post-processing; plotting supports inspection. The graph follows README claims and indexed imports/calls; some training, evaluation, and post-processing internals were not fully sampled, so unsupported links are omitted.
Main entry points: svg_loader.py Dataset loader · train.py Training loop · eval.py Evaluation runner · metrics.py Metrics
Components 19
Dataset and Loading
- House parser
floortrans/loaders/house.py - SVG geometry
floortrans/loaders/svg_utils.py - Dataset loader
floortrans/loaders/svg_loader.py - Augmentations
floortrans/loaders/augmentations.py - LMDB builder
create_lmdb.py - LMDB cache
data
Model and Training
- Training loop
train.py - Model selection
floortrans/models/__init__.py - Hourglass network
floortrans/models/hg_furukawa_original.py - Network blocks
floortrans/models/model_1427.py - Multi-task loss
floortrans/losses/uncertainty_loss.py
Prediction Processing
- Polygon post-processing
floortrans/post_prosessing.py
Evaluation and Outputs
- Evaluation runner
eval.py - Metrics
floortrans/metrics.py - Visualization
floortrans/plotting.py
Other components
- Dataset user
- SVG annotations
- Model developer
- Prediction user
Connections 21
- Dataset user → SVG annotations: provides plans
README.md - SVG annotations → House parser: parsed into labels
- House parser → SVG geometry: uses geometry
floortrans/loaders/house.py - Dataset loader → House parser: builds samples
floortrans/loaders/svg_loader.py - Dataset loader → Augmentations: applies transforms
floortrans/loaders/svg_loader.py - Dataset loader → LMDB cache: reads cached samples
floortrans/loaders/svg_loader.py - LMDB builder → LMDB cache: writes cache
README.md - LMDB builder → Dataset loader: uses dataset parsing
create_lmdb.py - Model developer → Training loop: starts training
README.md - Training loop → Dataset loader: loads samples
train.py - Training loop → Model selection: requests model
train.py - Model selection → Hourglass network: constructs network
floortrans/models/__init__.py - Hourglass network → Network blocks: uses architecture
floortrans/models/hg_furukawa_original.py - Training loop → Multi-task loss: optimizes loss
train.py - Prediction user → Evaluation runner: runs evaluation
README.md - Evaluation runner → Dataset loader: loads samples
eval.py - Evaluation runner → Model selection: requests model
eval.py - Evaluation runner → Metrics: evaluates predictions
eval.py - Metrics → Polygon post-processing: post-processes output
floortrans/metrics.py - Metrics → Visualization: uses visualization
floortrans/metrics.py - Prediction user → Polygon post-processing: receives vector output