Project · Code & Data

Source code, models, and data

Everything you need to reproduce the experiments — training scripts, evaluation pipelines, and dataset preparation utilities — lives in the GitHub repository linked below.

GitHub repository

The full source for the project — data preparation, model training, evaluation pipelines, and segment-level aggregation — is hosted on GitHub.

Open repository on GitHub

Trained model weights and the manual real-world segment images live as release assets on the same repo — see the v1.0 release.

Direct downloads

Code, datasets, trained weights, and the manual real-world segment images. Source and weights are hosted on the project's GitHub repo and v1.0 release; the two Mapillary datasets must be downloaded from Mapillary directly.

Source ZIP

Snapshot of the repository at the time of submission. Easiest way to grab everything in one go.

Download source (.zip)

Mapillary Traffic Sign Detection (MTSD)

The Mapillary Traffic Sign Detection dataset.

Download Data

Mapillary Vistas

The Mapillary Vistas dataset is a street-level imagery dataset used in the Parking Meter and Curb detection experiments.

Download Data

Trained sign detector

YOLOv8m parking-sign detector weights (50-epoch checkpoint, mAP@50 = 0.5487). Useful as a drop-in starting point.

Download weights (.pt)

Curb segmentation model

U-Net curb segmentation weights (best-Dice checkpoint, epoch 17, val Dice = 0.5184).

Download weights (.pt)

Manual segment images

The 30 manually collected real-world images (six segments × five views) used in the qualitative aggregation experiments.

Download images (.zip)

What's in the repository

High-level layout of QUASARS06/street-parking-presence-inference. Each top-level folder is a self-contained sub-project with its own scripts and outputs.

Top-level layout

street-parking-presence-inference/
├── mtsd_project/                # Parking-sign detection on MTSD
│   ├── scripts/                 #   Data prep, training, evaluation, viz
│   ├── dataset/                 #   YOLO-format binary parking-sign dataset (5-image samples)
│   ├── data/                    #   Processed images/labels (samples)
│   ├── outputs/                 #   Eval CSVs, threshold sweeps, category samples
│   ├── sample_images/           #   Manually collected qualitative images
│   └── parking_signs.yaml       #   YOLO data config
│
├── Mapillary_Vistas_Dataset/    # Curb segmentation + zero-shot meter eval
│   ├── scripts/                 #   Curb U-Net training, meter zero-shot eval
│   ├── curb_segmentation/       #   Mask metadata + samples
│   ├── training/, validation/, testing/   # 5-image dataset samples
│   ├── parking_meter_imgs/      #   Vistas parking-meter samples
│   └── outputs/                 #   Meter eval CSVs
│
├── manual_segments_dataset/     # 30-image real-world qualitative dataset
│   ├── segments/                #   6 segments × 5 views (raw images)
│   ├── outputs/                 #   Visualization outputs + scores CSVs
│   └── segments.csv             #   Index
│
└── parking_aggregation_project/ # Segment-level aggregation system
    ├── scripts/                 #   Synthetic-segment builder, aggregation, eval
    ├── metadata/                #   Per-image cue scores
    └── outputs/synthetic_segments/  # Synthetic benchmark + per-threshold results

Reproducing the main results

  1. Download MTSD and Mapillary Vistas v2.0 from their official sources. The repo only ships 5-image samples for each split.
  2. Build the binary parking-sign YOLO dataset and run training / evaluation from mtsd_project/scripts/.
  3. Train the curb segmentation U-Net and run the zero-shot parking-meter evaluation from Mapillary_Vistas_Dataset/scripts/.
  4. Build the synthetic pseudo-segment benchmark and run the segment-level aggregation evaluation from parking_aggregation_project/scripts/.
  5. Reproduce the annotated qualitative figures by running the visualization scripts against the manual_segments_dataset/ bundle (also available as segments.zip from the v1.0 release).

Environment

The experiments were primarily run on Kaggle (Tesla T4, P100) and a local Apple Silicon machine. Approximate environment:

  • Python 3.10+
  • PyTorch 2.x
  • Ultralytics YOLO (ultralytics package)
  • segmentation-models-pytorch for the curb U-Net
  • opencv-python, numpy, scikit-learn, matplotlib

Exact versions and the recommended environment are in the repo's requirements.txt / environment.yml.

If you have trouble reproducing a result or running a script, please open an issue on the GitHub repository and we'll take a look.