Note

This is a Hugging Face dataset. Learn how to load datasets from the Hub in the Hugging Face integration docs.

Hugging Face

Dataset Card for AVM (Around View Monitoring) Semantic Segmentation Dataset#

image/png

This repository provides a FiftyOne-compatible version of the AVM semantic segmentation dataset for autonomous parking systems, with enhanced metadata and visualization capabilities.

This is a FiftyOne dataset with 6763 samples.

Installation#

If you haven’t already, install FiftyOne:

pip install -U fiftyone

Usage#

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("harpreetsahota/AVM_Segmentation_train")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details#

Dataset Description#

The AVM dataset is a specialized computer vision dataset designed for training semantic segmentation models for autonomous parking systems. It contains bird’s-eye view images from around-view monitoring cameras with pixel-level annotations for parking space detection and obstacle avoidance.

Dataset Sources#

Uses#

Direct Use#

This dataset is designed for:

  • Autonomous Parking Systems: Training models to detect and navigate into parking spaces

  • Semantic Segmentation Research: Benchmarking segmentation algorithms on fisheye/bird’s-eye view images

  • Parking Space Detection: Identifying available vs occupied parking spots

  • Obstacle Detection: Recognizing curbs, pillars, walls, and other vehicles

  • 360° Surround View Systems: Enhancing camera-based parking assistance features

Out-of-Scope Use#

This dataset should NOT be used for:

  • Forward-facing autonomous driving (it’s specifically bird’s-eye view)

  • General object detection (annotations are polygon-based for segmentation)

  • High-speed navigation (designed for low-speed parking scenarios)

  • Pedestrian detection (pedestrians are not annotated)

Dataset Structure#

Overview#

  • Total Images: 6,763 (320 x 160 pixels)

  • Training Set: 4,057 images

  • Test Set: 2,706 images

  • Outdoor Images: 3,614

  • Indoor Images: 3,149

Semantic Classes#

The dataset contains 5 semantic classes with specific RGB color mappings:

Class

Description

RGB Color

Hex Color

0

Free Space (drivable area)

[0, 0, 255]

#0000FF (Blue)

1

Marker (parking lines)

[255, 255, 255]

#FFFFFF (White)

2

Vehicle (other cars)

[255, 0, 0]

#FF0000 (Red)

3

Other (curbs, pillars, walls)

[0, 255, 0]

#00FF00 (Green)

4

Ego Vehicle (camera car)

[0, 0, 0]

#000000 (Black)

FiftyOne Fields#

When parsed into FiftyOne, each sample includes:

Field

Type

Description

filepath

string

Path to the image file

split

string

“train” or “test”

sample_id

int

Unique identifier from filename

environment

Classification

“indoor” or “outdoor” (heuristic based on curb presence)

parking_type

Classification

“perpendicular” or “parallel”

slot_type

Classification

“closed”, “opened”, or “no_marker”

polygon_annotations

Polylines

Normalized polygon coordinates for each object

ground_truth

Segmentation

Pixel-level segmentation mask

classes_present

list

Classes present in the image

num_markers

int

Count of parking marker polygons

num_vehicles

int

Count of vehicle polygons

has_curb

bool

Whether curb is present

has_ego_vehicle

bool

Whether ego vehicle is annotated

Dataset Creation#

Curation Rationale#

The dataset was created to address the lack of bird’s-eye view datasets for autonomous parking systems. Most existing datasets focus on forward-facing cameras, but parking assistance requires a top-down perspective to accurately detect parking spaces and navigate safely.

Source Data#

Data Collection and Processing#

  • Camera Setup: Around View Monitoring (AVM) system with fisheye cameras

  • View Angle: Bird’s-eye view (top-down perspective)

  • Resolution: 320 x 160 pixels (optimized for embedded systems)

  • Environments: Real parking lots (both indoor parking garages and outdoor lots)

  • Conditions: Various lighting conditions, weather (sunny, cloudy, rainy)

Who are the source data producers?#

The original dataset was produced by researchers developing autonomous parking systems, likely in an academic or industrial research setting.

Annotations#

Annotation Process#

  1. Polygon Annotation: Each object is annotated with precise polygon boundaries in YAML format

  2. Semantic Masks: Ground truth masks are generated from polygon annotations

  3. Multi-polygon Support: Multiple instances of the same class are supported (e.g., multiple vehicles)

  4. Coordinate System: Polygons use image coordinates (0-319 x 0-159)

Who are the annotators?#

Information about specific annotators is not provided in the original dataset documentation.

Personal and Sensitive Information#

The dataset contains images from parking lots but does not include:

  • License plate information (resolution too low)

  • Personally identifiable information

  • Pedestrian annotations

  • Location-specific information

Bias, Risks, and Limitations#

Known Limitations#

  1. Limited Resolution: 320x160 pixels may not capture fine details

  2. Geographic Bias: Dataset may be from specific geographic regions

  3. Weather Conditions: Limited representation of extreme weather

  4. Vehicle Types: May not include all vehicle types (trucks, motorcycles, etc.)

  5. Parking Styles: Primarily perpendicular and parallel parking

Technical Challenges#

  • Indoor Reflections: Reflected lights can be mistaken for parking markers

  • Fisheye Distortion: Bird’s-eye view introduces geometric distortions

  • Class Imbalance: Some classes (like curbs) appear less frequently

Recommendations#

  1. Augmentation: Apply data augmentation to improve model robustness

  2. Validation: Test models on diverse parking environments not in the dataset

  3. Resolution: Consider upscaling techniques if higher resolution is needed

  4. Edge Cases: Be aware that the dataset may not cover all parking scenarios

Exploring the Dataset#

# View class distribution
print(dataset.count_values("classes_present"))

# Filter indoor vs outdoor
indoor = dataset.match(F("environment.label") == "indoor")
outdoor = dataset.match(F("environment.label") == "outdoor")

# Samples with multiple vehicles
multi_vehicle = dataset.match(F("num_vehicles") > 2)

Citation#

BibTeX#

@dataset{avm_dataset,
  title={AVM (Around View Monitoring) System Datasets for Auto Parking},
  author={Chulhoon Jang and others},
  year={2020},
  url={https://github.com/ChulhoonJang/avm_dataset}
}

APA#

Jang, C., et al. (2020). AVM (Around View Monitoring) System Datasets for Auto Parking. GitHub. https://github.com/ChulhoonJang/avm_dataset

More Information#

Dataset Statistics#

  • Average polygons per class:

    • Ego vehicle: 1.0 polygons (fixed position)

    • Markers: 2.6 polygons per image

    • Vehicles: 2.1 polygons per image

    • Curbs: 1.4 polygons per image (when present)

Dataset Card Authors#

  • FiftyOne Integration: Harpreet Sahota (Voxel51)

  • Original Dataset: Chulhoon Jang and team

Dataset Card Contact#

Acknowledgments#

Thanks to the original dataset creators for making this valuable resource available to the research community. The FiftyOne integration enhances the dataset’s usability for modern computer vision workflows.