Point-supervised remote sensing segmentation

Personal short test project implementing partial cross-entropy (pCE) and partial focal cross-entropy (pfCE) on simulated point labels, trained on LoveDA (real remote sensing) or a synthetic benchmark for fast checks.

What is included

Item	Description
src/losses.py	Masked pCE and pfCE (normalized by number of labeled pixels).
src/data.py	LoveDA loader (TorchGeo), random/balanced point simulation, synthetic dataset.
src/model.py	U-Net + ResNet-34 (ImageNet weights) via segmentation_models_pytorch.
train.py	Training and full-mask validation mIoU (LoveDA label 0 = no-data excluded from IoU mean).
run_experiments.py	Experiment grid → experiments/logs/results.csv.
tests/test_losses.py	Sanity checks for the loss implementations.
TECHNICAL_REPORT.md	Method, experiments (purpose / hypothesis / procedure), and results.

Requirements

• Python 3.10+ (3.11 recommended)
• Disk: a few MB for code; ~4 GB+ per split if you download LoveDA train/val
• GPU: optional; CPU runs work for small epochs and synthetic data

Setup (virtual environment)

Do not install dependencies system-wide.

cd /path/to/TestProject
python3 -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
python -m pip install --upgrade pip
pip install -r requirements.txt

For GPU PyTorch, install the matching torch / torchvision wheels from pytorch.org (then install the rest of requirements.txt if needed).

Quick verification

source .venv/bin/activate   # or: .venv/bin/python tests/test_losses.py
python tests/test_losses.py

Expected: ok
If your shell does not load the venv (some CI setups), call the interpreter explicitly: .venv/bin/python tests/test_losses.py.

Training

Run from the project root (the repo root that contains src/ and train.py).

Synthetic (no download; good for CI and smoke tests)

python train.py --dataset synthetic --epochs 8 --batch-size 8 --crop-size 256 \
  --num-points 50 --loss focal --sampling random --seed 42

LoveDA (real remote sensing)

First run downloads Zenodo archives into --data-root (large). Use caps while developing:

python train.py --dataset loveda --data-root ./data/loveda --download \
  --epochs 10 --batch-size 4 --crop-size 512 \
  --num-points 50 --loss focal --sampling random \
  --max-train-samples 400 --max-val-samples 150 --seed 42

Note on mask PNGs: segmentation masks use small integer class IDs (0–7). Many viewers show them as almost black; that is normal. See TECHNICAL_REPORT.md (FAQ).

Experiments (CSV)

Runs Experiment A (point density × sampling, pfCE) and Experiment B (pCE vs pfCE at fixed settings):

python run_experiments.py --dataset synthetic --epochs 3 --batch-size 8 --crop-size 256 \
  --max-train-samples 200 --max-val-samples 80 --seed 42

LoveDA (after data is present):

python run_experiments.py --dataset loveda --data-root ./data/loveda --epochs 5 \
  --max-train-samples 500 --max-val-samples 200
# Add --download only if ./data/loveda is not populated yet.

Output: experiments/logs/results.csv

Main CLI options (train.py)

Option	Meaning
--dataset	synthetic or loveda
--download	Download LoveDA into --data-root (huge)
--num-points	Simulated clicks per crop (N)
--sampling	random (PDF baseline) or balanced (experiment)
--loss	focal (pfCE) or ce (pCE)
--epochs, --batch-size, --lr, --crop-size	Training knobs
--max-train-samples, --max-val-samples	Subset caps for speed

Creating a project zip

Exclude the venv and any downloaded data. From the project root:

./scripts/package_submission.sh ./submission.zip

Or manually zip the repository without .venv/, data/, and __pycache__/.

The produced archive should include: src/, train.py, run_experiments.py, requirements.txt, tests/, experiments/logs/results.csv (after you run experiments), TECHNICAL_REPORT.md, and this README.md.

Citation (LoveDA)

If you use LoveDA, cite:

Wang et al., LoveDA: A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation, arXiv:2110.08733.

License

This repository is a personal short test project. LoveDA has its own license; follow the LoveDA repository terms when using the data.

pfce-point-segmentation