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

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.

Curated by: Chulhoon Jang and team at original repository
FiftyOne Integration by: Harpreet Sahota (Voxel51)
License: Please refer to the original dataset repository for license information (which currently has no License)

Dataset Sources

Original Repository: https://github.com/ChulhoonJang/avm_dataset

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

Polygon Annotation: Each object is annotated with precise polygon boundaries in YAML format
Semantic Masks: Ground truth masks are generated from polygon annotations
Multi-polygon Support: Multiple instances of the same class are supported (e.g., multiple vehicles)
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

Limited Resolution: 320x160 pixels may not capture fine details
Geographic Bias: Dataset may be from specific geographic regions
Weather Conditions: Limited representation of extreme weather
Vehicle Types: May not include all vehicle types (trucks, motorcycles, etc.)
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

Augmentation: Apply data augmentation to improve model robustness
Validation: Test models on diverse parking environments not in the dataset
Resolution: Consider upscaling techniques if higher resolution is needed
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

Related Resources

Original Dataset Repository
FiftyOne Documentation
Implementation code for semantic segmentation models (link in original repo)

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

Original dataset: See original repository

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.

image imagewidth (px) 320 1.71k