Compute-Constrained Road Damage Detection

Reproducible CPU-only object detection on the multinational RDD2022 benchmark, with sequence-aware splitting, explicit negative-image preservation, equal-domain training, and domain-sliced COCO evaluation.

The frozen YOLO11n run uses 2,800 training images (400 per acquisition domain), 10 epochs, 320-by-320 inputs, and no GPU. On the held-out 3,703-image internal test split it obtains 3.26% COCO mAP, 8.51% AP50, and 43.4 images/s. The result is intended as an auditable lower-compute reference, not a state-of-the-art claim.

Paper

• Compiled IEEE paper
• LaTeX source
• Results provenance
• Permanent Zenodo archive

The paper is written in English using the official IEEEtran conference class. Build it with:

cd paper
make

Installation

The published run used Python 3.14, while CI verifies the lightweight pipeline on Python 3.12. A virtual environment is strongly recommended:

git clone https://github.com/Mathweuzz/rdd2022-cpu-baseline.git
cd rdd2022-cpu-baseline
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[experiments]"

Install manifests pin patched compatible versions. The exact historical versions used to produce the published metrics remain recorded in paper/RESULTS_PROVENANCE.md; they are provenance, not a recommendation to install packages with known advisories.

Reproduce the complete pipeline

The dataset is not redistributed by this repository. Download RDD2022 from its Figshare record, then run:

make download   # approximately 13 GB, with Figshare MD5 verification
make eda        # reads the seven domain ZIP files directly
make prepare    # sequence-blocked split plus COCO/YOLO export
make validate   # structural, geometry, and leakage checks
make protocol   # frozen 2,800-image equal-domain training lists

The commands intentionally remain separate so every generated artifact can be audited before the next stage. Run make help for all entry points.

Reproduce training and evaluation

The following command implements the exact frozen CPU defaults from the paper:

make train

After training, evaluate the final-epoch checkpoint with the low-memory path:

OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 \
python experiments/evaluate_yolo_checkpoint.py \
  --model outputs/yolo11n_joint_cpu_final_seed2026/weights/last.pt \
  --split test \
  --output outputs/yolo11n_joint_cpu_final_seed2026/test_evaluation \
  --imgsz 320 --batch 1 --workers 0 --threads 1

To try the released checkpoint on one image or a directory:

gh release download v0.1.0 --pattern 'yolo11n-rdd2022-cpu-seed2026.*' --dir models
python experiments/predict.py \
  --model models/yolo11n-rdd2022-cpu-seed2026.safetensors \
  --config models/yolo11n-rdd2022-cpu-seed2026.yaml \
  --source path/to/image-or-directory

Experiment entry points and the frozen CPU protocol are documented in experiments/README.md and EXPERIMENT_PROTOCOL.md. The end-to-end artifact map is in docs/REPRODUCIBILITY.md.

Repository contents

• artifacts/: compact, machine-readable frozen metrics and training history;
• experiments/: training, evaluation, diagnostics, and experiment log;
• paper/: IEEE manuscript, bibliography, template provenance, and figure;
• tests/: lightweight integrity and determinism checks;
• root Python scripts: download, cleaning, validation, and EDA.

Datasets, dependency environments, checkpoints, raw predictions, and generated experiment directories are excluded from Git history. Artifact SHA-256 values are recorded so externally hosted weights and predictions can be verified.

Verification

make test
make paper

GitHub Actions repeats the lightweight tests and builds the paper from a clean checkout. The full 28.5-minute training run is intentionally not repeated in CI.

Author

Mateus Gomes de Araújo, Computer Engineering, University of Brasília (UnB). Contact: mathweuzz@gmail.com.

Citation

The versioned software archive is available under DOI 10.5281/zenodo.22285376. Use the concept DOI 10.5281/zenodo.22285375 to cite the project independently of a specific release. Machine-readable citation metadata is provided in CITATION.cff.

License

Original software in this repository is released under the MIT License. The manuscript and original figures remain copyright © 2026 Mateus Gomes de Araújo. The vendored IEEEtran files retain their upstream license and copyright notices. Ultralytics and the derived checkpoint are subject to AGPL-3.0 upstream terms. RDD2022 is not redistributed and remains subject to its source terms. See third-party notices for scope details.