LawnForge-NET: Lawn Segmentation From Dense Point Clouds

Production-oriented pipeline for extracting lawn polygons and per-point lawn labels from dense outdoor point clouds.

Project Goals

• Segment lawn areas robustly in dense street scans.
• Support very large scenes (millions of points).
• Export GIS-ready vector outputs and point cloud predictions with preserved CRS.
• Provide a GUI workflow for non-code operation.

Core Approach

The current pipeline is a hybrid method:

1. CSF (Cloth Simulation Filter) removes non-ground points.
2. A 2.5D U-Net performs lawn vs non-lawn segmentation on ground-only data.
3. Predictions are projected back to points and exported as:

• point_predictions.laz
• lawn_polygons.gpkg

Why this approach:

• Better robustness against trees/buildings than plain raster segmentation.
• Lower memory footprint than full 3D sparse models.
• Strong GIS compatibility.

Repository Layout

• configs/default.yaml: default pipeline config.
• scripts/prepare_dataset.py: tile generation for training.
• scripts/train_stage2.py: stage-2 model training.
• scripts/infer_scene.py: single-scene inference.
• scripts/gui_app.py: full GUI launcher.
• src/csf_filter.py: CSF ground filtering wrapper.
• src/raster.py: 2.5D feature rasterization and tiling.
• src/models/unet2d.py: 2D U-Net segmentation model.
• src/train_utils.py: training loop and metrics.
• src/infer_utils.py: tiled inference and point projection.
• src/io_utils.py: scene reading/writing, LAZ export.
• src/vectorize_utils.py: polygonization and GPKG export.
• run.bat: one-click setup and GUI start for Windows.

System Requirements

See full details in docs/SYSTEM_REQUIREMENTS.md.

Baseline target:

• Python 3.11
• CUDA 11.8 (PyTorch cu118 build)
• NVIDIA GPU with 6 GB+ VRAM recommended
• Windows 10/11 x64

Quick Start (Windows)

1. Place data in expected folders:

• data/raw/train
• data/raw/val

2. Launch:

run.bat

run.bat will:

• create/update temporary venv (.venv_temp_gui)
• install dependencies
• install PyTorch CUDA 11.8 build
• start GUI

Manual CLI Workflow

1) Prepare dataset

python scripts/prepare_dataset.py --config configs/default.yaml

2) Train model

python scripts/train_stage2.py --config configs/default.yaml --weights-out artifacts/models/stage2_lawn_vs_nonlawn.pt

3) Inference (minimal outputs)

python scripts/infer_scene.py --config configs/default.yaml --scene data/raw/val/scene_001.laz --out-dir outputs/scene_001 --weights artifacts/models/stage2_lawn_vs_nonlawn.pt --minimal-outputs

Outputs

Minimal mode (--minimal-outputs):

• point_predictions.laz
• lawn_polygons.gpkg

Full mode:

• point_predictions.laz
• lawn_polygons.gpkg
• lawn_polygons.geojson
• lawn_polygons.shp
• class_raster.tif
• lawn_mask.tif
• csf_ground_mask.npz
• point_labels.npz

Data Requirements

Training scenes require:

• x, y, z, intensity, label

Inference scenes require:

• x, y, z, intensity
• label/classification optional

Supported formats:

• .laz/.las
• .npz
• .csv

Coordinate Reference System (CRS)

• CRS is read from input LAS/LAZ when available.
• LAZ prediction output preserves original dimensions and CRS metadata.
• GPKG polygons are written with CRS WKT.

GUI Capabilities

• Full settings editing in UI (self-contained, no manual YAML required).
• Batch inference over multiple files.
• Per-file output subfolders.
• Weights file selection for train/inference.
• Detailed live logs.
• Status checks:
• Torch
• CUDA
• Model checkpoint load
• RU/EN language switch.

Documentation Index

• docs/SYSTEM_REQUIREMENTS.md
• docs/METHODS.md
• docs/INSTALLATION.md
• docs/USAGE.md
• docs/DATA_FORMAT.md
• docs/TRAINING_DATA_PREP.md

Current Scope

Current model predicts lawn-oriented classes only (lawn vs non-lawn ground after CSF filtering). This repository is structured for extension to more vegetation classes.