Datasets:

DSBD-Research
/

MetaPKLot-Dataset

Tasks:

Modalities:

Image

Formats:

Size:

Tags:

occupancy-recognition

Libraries:

Datasets

License:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Viewer

Auto-converted to Parquet Duplicate

Split (1)

train · 16.5k rows

Search is not available for this dataset

image imagewidth (px) 1k 1.28k

End of preview. Expand in Data Studio

MetaPKLot

A Large-Scale Benchmark for Vision-Based Parking Lot Management

2,265,974 labeled samples · 1,366,185 new annotations · 3 research challenges · COCO-style annotations

MetaPKLot is a large-scale, harmonized dataset designed for research on vision-based parking lot management.

It extends and standardizes three existing parking datasets:

PKLot
CNRPark-EXT
PLds

MetaPKLot introduces new annotations, revises existing parking-space annotations, standardizes their representation, and provides a unified benchmark for evaluating parking-management systems under realistic cross-dataset conditions.

The dataset supports three main challenges:

Parking Spot Occupancy Recognition
Parking Dwell Time Estimation
Parking Spot Extraction

Dataset Summary

MetaPKLot extends PKLot, CNRPark-EXT, and PLds with a common annotation structure suitable for modern computer vision research.

The dataset adds 1,366,185 new annotations and revises approximately 900,000 existing annotations, resulting in a total of:

2,265,974 labeled samples

The new annotations include:

570,326 new labeled parking spots
795,859 vehicle segmentation polygons
More than 32,000 manually identified vehicles
Standardized parking-space polygons
Parking-space occupancy labels
Vehicle instance segmentation masks
Persistent vehicle identifiers
Image timestamps
Information required for dwell-time estimation
Bounding boxes
Rotated rectangles for parking-space crops
COCO-style JSON annotations

The same physical vehicle can retain the same car_id across multiple images, enabling temporal analysis and parking dwell-time estimation.

Dataset at a Glance

Property	MetaPKLot
Source datasets	PKLot, CNRPark-EXT, PLds
Total labeled samples	2,265,974
New annotations	1,366,185
Revised annotations	Approximately 900,000
New labeled parking spots	570,326
Vehicle segmentation polygons	795,859
Manually identified vehicles	More than 32,000
Annotation format	COCO-style JSON
Parking occupancy labels	Empty / Occupied
Vehicle identities	Persistent `car_id`
Main challenges	Occupancy Recognition, Dwell Time Estimation, Parking Spot Extraction

Research Challenges

MetaPKLot defines three standardized research challenges for vision-based parking management.

1. Parking Spot Occupancy Recognition

The goal is to determine whether a predefined parking spot is:

Empty
Occupied

The parking-space position is assumed to be known beforehand.

This challenge is intended to evaluate models capable of generalizing across different parking environments and camera configurations.

The primary recommended classification metric is Macro F1.

Computational efficiency is also relevant, particularly because parking-management systems may be deployed on edge devices such as smart cameras.

2. Parking Dwell Time Estimation

The goal is to estimate how long a vehicle remains parked.

MetaPKLot provides persistent vehicle identifiers through the car_id field. When the same physical vehicle appears across multiple images, its identifier can be used to associate observations over time.

Together with image timestamps and parking-space annotations, these identifiers enable the development and evaluation of dwell-time estimation methods.

Potential applications include:

Parking billing
Parking-duration monitoring
Long-term vehicle detection
Parking behavior analysis
Smart parking management

3. Parking Spot Extraction

The goal is to automatically discover and delineate parking-space locations directly from parking-lot images.

This is particularly useful when predefined parking-space templates are unavailable.

MetaPKLot provides:

Parking-space polygons
Vehicle segmentation masks
Temporal observations
Standardized annotations

These resources enable approaches based on vehicle occurrence, segmentation, spatial accumulation, geometric analysis, and other computer vision techniques.

Experimental Protocol

Golden Rule

All experiments should follow a cross-dataset evaluation using a leave-one-dataset-out strategy.

The objective is to evaluate whether a method can generalize to a parking environment from which no labeled training samples are available.

Since MetaPKLot currently contains three source datasets, the recommended configurations are:

Configuration 1

Training / Validation:
PKLot + CNRPark-EXT

Testing:
PLds

Configuration 2

Training / Validation:
PKLot + PLds

Testing:
CNRPark-EXT

Configuration 3

Training / Validation:
CNRPark-EXT + PLds

Testing:
PKLot

The test dataset must remain isolated from training.

Researchers may define their own training and validation split inside the two training datasets, but labeled samples from the test dataset should not be used for model training.

This protocol is designed to measure generalization across:

Unseen parking lots
Different camera viewpoints
Different image resolutions
Different lighting conditions
Different weather conditions
Different parking layouts

Dataset Organization

The MetaPKLot repository contains images, annotations, dataset metadata, correction lists, and conversion tools.

A simplified view of the repository is:

MetaPKLot-Dataset/
├── CNRPark-EXT/
├── PKLot/
├── annotations/
│   ├── cropped/
│   │   ├── cars/
│   │   └── spots/
│   └── original/
│       ├── cars/
│       └── spots/
├── non_representative_days/
├── non_working_days/
├── removed_images/
├── tools/
├── readme_images/
└── To-All-Who-Contributed.txt

The canonical project repository is:

DSBD-Research/MetaPKLot-Dataset

Reference implementations and baseline models for the proposed challenges are maintained separately in:

DSBD-Research/MetaPKLot-Challenges

Annotation Format

The annotation files follow the general structure of the Microsoft COCO format, with additional fields required by parking-management applications.

The annotations are primarily divided into two types:

Cars
Spots

Cars JSON

The Cars JSON files describe annotated vehicles.

They may contain:

Vehicle segmentation polygons
Bounding boxes
Object area
Image identifiers
Persistent car_id
Dataset information
Camera/subset information
Date
Time
Climate information

If an image does not contain an annotated vehicle, it may not appear in a Cars JSON file.

The same physical vehicle appearing across different images receives the same car_id when the annotation process identifies it as the same vehicle.

This persistent identifier is particularly important for the Parking Dwell Time Estimation challenge.

A simplified structure is shown below:

{
  "categories": [
    {
      "supercategory": "vehicle",
      "name": "car",
      "id": 0
    }
  ],
  "climates": [
    {
      "id": 1,
      "name": "overcast"
    },
    {
      "id": 2,
      "name": "sunny"
    },
    {
      "id": 3,
      "name": "rainy"
    },
    {
      "id": 4,
      "name": "snowy"
    },
    {
      "id": 5,
      "name": "no_rain"
    },
    {
      "id": 6,
      "name": "undefined"
    }
  ],
  "images": [
    {
      "id": 0,
      "file_name": "UFPR04/Sunny/2012-12-07/2012-12-07_20_17_28.jpg",
      "width": 1280,
      "height": 720,
      "annotationsRectangle": [
        "upper_left_x",
        "upper_left_y",
        "width",
        "height"
      ],
      "date": [
        2012,
        12,
        7
      ],
      "time": [
        20,
        17,
        28
      ],
      "climate": 2,
      "dataset": "pklot",
      "subset": "ufpr04"
    }
  ],
  "annotations": [
    {
      "id": 0,
      "segmentation": [
        [
          "x1",
          "y1",
          "x2",
          "y2",
          "...",
          "xn",
          "yn"
        ]
      ],
      "area": "area_in_pixels",
      "bbox": [
        "upper_left_x",
        "upper_left_y",
        "width",
        "height"
      ],
      "image_id": 0,
      "car_id": 500,
      "category_id": 0
    }
  ]
}

Spots JSON

The Spots JSON files describe parking-space annotations.

They may contain:

Parking-space segmentation polygons
Occupancy status
Associated car_id
Bounding boxes
Rotated rectangles
Image metadata
Dataset metadata
Date
Time
Climate information

Parking-space categories are represented as:

category_id = 0  -> empty
category_id = 1  -> occupied

The car_id field has additional special values:

car_id = -1  -> no vehicle in the parking spot
car_id = -2  -> parking spot occupied by something other than a car

When a parking spot contains an annotated vehicle, its car_id corresponds to the same identifier used for that vehicle in the Cars JSON.

A simplified structure is shown below:

{
  "categories": [
    {
      "id": 0,
      "name": "empty"
    },
    {
      "id": 1,
      "name": "occupied"
    }
  ],
  "climates": [
    {
      "id": 1,
      "name": "overcast"
    },
    {
      "id": 2,
      "name": "sunny"
    },
    {
      "id": 3,
      "name": "rainy"
    },
    {
      "id": 4,
      "name": "snowy"
    },
    {
      "id": 5,
      "name": "no_rain"
    },
    {
      "id": 6,
      "name": "undefined"
    }
  ],
  "images": [
    {
      "id": 0,
      "file_name": "UFPR04/Sunny/2012-12-07/2012-12-07_20_17_28.jpg",
      "width": 1280,
      "height": 720,
      "annotationsRectangle": [
        "upper_left_x",
        "upper_left_y",
        "width",
        "height"
      ],
      "date": [
        2012,
        12,
        7
      ],
      "time": [
        20,
        17,
        28
      ],
      "climate": 2,
      "dataset": "pklot",
      "subset": "ufpr04"
    }
  ],
  "annotations": [
    {
      "id": 0,
      "segmentation": [
        [
          "x1",
          "y1",
          "x2",
          "y2",
          "...",
          "xn",
          "yn"
        ]
      ],
      "bbox": [
        "upper_left_x",
        "upper_left_y",
        "width",
        "height"
      ],
      "image_id": 0,
      "car_id": 500,
      "category_id": 1,
      "best_rotated_rect": [
        [
          "center_x",
          "center_y"
        ],
        [
          "width",
          "height"
        ],
        "angle"
      ]
    }
  ]
}

`annotationsRectangle`

Some images contain an annotationsRectangle field.

This rectangle defines the region in which the annotation team ensured that the relevant objects were consistently annotated.

The general format is:

[
  upper_left_x,
  upper_left_y,
  width,
  height
]

Only objects located inside this annotation region should be assumed to have complete annotations.

Objects outside the region should not automatically be considered unlabeled negatives.

This is especially important when:

Training object detectors
Training instance segmentation models
Computing evaluation metrics
Creating parking-space crops
Evaluating false-positive detections

Segmentation Polygons

Vehicle segmentation masks are represented using polygons.

Each vehicle polygon contains at least three points and may use additional vertices when necessary to represent the shape of the vehicle.

Parking-space positions are also standardized using polygon representations.

MetaPKLot harmonizes the different annotation styles originally used by PKLot, CNRPark-EXT, and PLds into a more consistent representation.

Bounding Boxes

Bounding boxes use the general COCO convention:

[
  upper_left_x,
  upper_left_y,
  width,
  height
]

They can be used for tasks such as:

Vehicle detection
Visualization
Image cropping
Region extraction
Preprocessing

Rotated Parking-Space Rectangles

Parking-space annotations may include:

best_rotated_rect

The general structure is:

[
  [center_x, center_y],
  [width, height],
  angle
]

This field can be used to extract normalized parking-space crops while accounting for the orientation of parking spaces in the image.

File Paths

File paths stored in the JSON annotations follow the POSIX convention.

Example:

root_directory/subdirectory/filename.extension

For example:

UFPR04/Sunny/2012-12-07/2012-12-07_20_17_28.jpg

Users working on Windows may need to convert path separators when processing these paths using operating-system-specific APIs.

Climate Metadata

Images may contain a climate identifier.

The annotation structure includes the following climate categories:

ID	Climate
1	overcast
2	sunny
3	rainy
4	snowy
5	no_rain
6	undefined

Not every source dataset necessarily contains samples for every climate category.

Source Datasets

PKLot

PKLot is one of the source datasets incorporated into MetaPKLot.

MetaPKLot includes revised and additional annotations for PKLot and provides the corresponding images according to the licensing conditions of the original dataset.

The PKLot dataset is licensed under:

Creative Commons Attribution 4.0 International (CC BY 4.0)

CNRPark-EXT

CNRPark-EXT is another source dataset incorporated into MetaPKLot.

MetaPKLot includes revised and additional annotations together with the corresponding images according to the original dataset licensing conditions.

CNRPark-EXT is licensed under:

Open Data Commons Open Database License (ODbL) v1.0

PLds

MetaPKLot also provides annotations corresponding to the PLds dataset.

However:

The original PLds images are not redistributed as part of MetaPKLot.

Researchers who want to use the PLds portion of MetaPKLot must obtain the original PLds images from the original dataset authors.

Additional information and tools for integrating PLds are available in:

tools/PLds/

in the canonical MetaPKLot repository.

Data Availability

The MetaPKLot repository provides:

PKLot images and MetaPKLot annotations
CNRPark-EXT images and MetaPKLot annotations
PLds MetaPKLot annotations
Dataset correction information
Lists of excluded images
Experimental metadata
Conversion and support tools

The original PLds images must be obtained separately.

Dataset Corrections and Curation

MetaPKLot includes corrections and quality-control information for the original datasets.

For example, known PKLot directory/date corrections include:

UFPR05/Sunny/2013-14-16

corrected to:

UFPR05/Sunny/2013-04-16

and:

PUCPR/Cloudy/2012-09-16

corrected to:

PUCPR/Cloudy/2012-10-16

Information about removed images is available in:

removed_images/

This includes, among other cases:

Corrupted PKLot images
CNRPark-EXT images with unsuitable lighting conditions
PLds images removed because of repeated timestamps
PLds images affected by unexpected zoom

Additional experimental metadata is available in:

non_representative_days/

and:

non_working_days/

Researchers are encouraged to consult these files when reproducing experiments.

Annotation Quality

MetaPKLot was created with an extensive manual annotation process.

The associated publication reports quality-control experiments using independently annotated subsets from all three source datasets.

Reported agreement included:

99.62% agreement for parking-space occupancy labels
Average IoU of 0.85 for parking-space rotated rectangles
Average IoU of 0.90 for vehicle polygons
99.64% agreement for vehicle identity annotations

Given the scale of the dataset, minor annotation errors may still exist.

The public repository allows corrections and improvements to be incorporated over time.

Recommended Uses

MetaPKLot is intended for research involving:

Parking spot occupancy recognition
Parking dwell time estimation
Parking spot extraction
Vehicle instance segmentation
Vehicle detection
Cross-domain computer vision
Domain adaptation
Domain generalization
Transfer learning
Smart-city applications
Edge computer vision
Parking analytics
Temporal vehicle association

Limitations

Users should consider the following limitations:

MetaPKLot combines existing parking datasets rather than covering every possible parking environment.
Camera viewpoints and parking layouts are constrained by the source datasets.
Some objects outside annotationsRectangle may not be annotated.
PLds images are not redistributed and must be obtained separately.
Despite extensive quality-control procedures, annotation errors may remain.
Real-world deployment may involve parking environments, vehicle types, weather conditions, and camera configurations that are not represented in the dataset.

Researchers should therefore avoid assuming that performance on MetaPKLot directly represents performance in every real-world parking environment.

Cross-dataset evaluation is strongly recommended.

Accessing the Dataset on Hugging Face

The Hugging Face dataset repository identifier is:

DSBD-Research/MetaPKLot-Dataset

The repository can be downloaded with the Hugging Face CLI:

hf download DSBD-Research/MetaPKLot-Dataset \
  --repo-type dataset \
  --local-dir MetaPKLot-Dataset

MetaPKLot is primarily distributed as image files, annotation JSON files, and supporting metadata.

Depending on the intended experiment, users may need to build a custom data-loading pipeline using the JSON annotations described above.

Canonical Repository

The canonical source-code and dataset repository is:

DSBD-Research/MetaPKLot-Dataset

This repository contains:

Dataset documentation
Annotation files
Images that can be redistributed
Dataset corrections
PLds integration tools
Removed-image lists
Experimental metadata
Supporting utilities

For the most recent project changes, researchers should consult the canonical repository.

Baselines and Models

Reference implementations and trained baseline models associated with the MetaPKLot challenges are available in the companion project:

DSBD-Research/MetaPKLot-Challenges

The baseline project covers the three main challenges:

classification/
dwell_time/
parking_spot_extraction/

These implementations are intended to support reproducibility and provide reference points for future research.

Licensing

Important

MetaPKLot combines resources originating from datasets with different licensing conditions.

Therefore, a single upstream license should not be assumed to apply to every file distributed as part of MetaPKLot.

The Hugging Face metadata uses:

license: other

to reflect these mixed licensing conditions.

PKLot

PKLot is distributed under:

Creative Commons Attribution 4.0 International (CC BY 4.0)

CNRPark-EXT

CNRPark-EXT is distributed under:

Open Data Commons Open Database License (ODbL) v1.0

PLds

The PLds images are not redistributed with MetaPKLot.

Users must obtain the images from the original PLds authors and comply with the applicable terms established by the original dataset providers.

Researchers should review the licensing conditions of each source dataset before redistributing files or creating derivative releases.

Acknowledgements

MetaPKLot builds upon the work of the researchers who created and released:

PKLot
CNRPark-EXT
PLds

We gratefully acknowledge their contributions to vision-based parking management research.

We also thank everyone who contributed to the annotation, validation, development, and maintenance of MetaPKLot.

The canonical repository contains an additional acknowledgement file:

To-All-Who-Contributed.txt

The research associated with MetaPKLot was supported by the Brazilian National Council for Scientific and Technological Development (CNPq) under Grant:

405511/2022-1

The Article Processing Charge of the associated publication was funded by CAPES — Brazil.

Associated Publication

MetaPKLot is described in:

Paulo Lisboa de Almeida, Jeovane Honorio Alves, Luan Marko Kujavski, Paulo Luza Alves, and Luiz Eduardo Oliveira.

MetaPKLot – new challenges and protocols for vision-based parking lot management.

Neural Computing and Applications, Volume 38, Article 704, 2026.

Published: 29 August 2026

DOI:

10.1007/s00521-026-12398-0

Citation

If MetaPKLot contributes to your research, please cite the associated publication.

@article{deAlmeida2026MetaPKLot,
  author = {de Almeida, Paulo Lisboa and
            Alves, Jeovane Honorio and
            Kujavski, Luan Marko and
            Alves, Paulo Luza and
            Oliveira, Luiz Eduardo},
  title = {MetaPKLot -- new challenges and protocols for vision-based parking lot management},
  journal = {Neural Computing and Applications},
  volume = {38},
  article = {704},
  year = {2026},
  doi = {10.1007/s00521-026-12398-0}
}

Contact and Contributions

For dataset issues, corrections, or improvements, please use the issue tracker in the canonical repository:

DSBD-Research/MetaPKLot-Dataset

Given the scale of MetaPKLot, community reports of annotation inconsistencies and dataset issues are welcome.

Reference

MetaPKLot — new challenges and protocols for vision-based parking lot management

Neural Computing and Applications, 2026

DOI: 10.1007/s00521-026-12398-0

Canonical dataset repository:

DSBD-Research/MetaPKLot-Dataset

Baseline implementations and models:

DSBD-Research/MetaPKLot-Challenges

Downloads last month: 408