Voxel51/DCVAI-Challenge-Public-Eval-Set

Dataset Card for Public Evaluation set for the Data Centric Visual AI Challenge

This is the public evaluation set for the Data Centric Visual AI Challenge.

This is a FiftyOne dataset with 1720 samples.

These images are a subset from Open Images v6 and Open Images v7 .

Full information about the dataset can be found here.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = fouh.load_from_hub("Voxel51/DCVAI-Challenge-Public-Eval-Set")

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

Dataset Overview

Welcome to our Data Curation Challenge for Object Detection! This page provides essential information about the dataset you'll be working with.

Dataset Description

Our dataset is a custom collection of images specifically curated for this competition. It focuses on everyday objects, people, and transportation, with a particular emphasis on clothing and accessories.

Key details:

• Total images: 1,720
• Image format: JPEG
• Annotation format: FiftyOne detection format

Object Classes

The dataset includes 43 object classes across several categories:

1. Clothing and Accessories: Footwear, Jeans, Trousers, Shorts, Glasses, Sunglasses, High heels, Boot, Sandal

2. People: Man, Woman, Girl, Boy

3. Transportation: Airplane, Car, Truck, Van, Ambulance, Helicopter, Motorcycle, Bicycle, Unicycle, Bus, Taxi

4. Electronics: Television, Computer monitor, Laptop, Mobile phone, Corded phone

5. Sports Equipment: Tennis ball, Tennis racket, Table tennis racket, Golf ball, Ball, Rugby ball, Football, Kite, Volleyball (Ball), Balloon

6. Food: Hamburger, Sandwich, Submarine sandwich, Hot dog

Annotations

Each image in the dataset comes with detailed annotations in FiftyOne detection format. A typical annotation looks like this:

<Detection: {
    'id': '66a037ceef34f40a421a9810',
    'attributes': {},
    'tags': [],
    'label': 'Jeans',
    'bounding_box': [0.446875, 0.36773, 0.16562500000000002, 0.321763],
    'mask': None,
    'confidence': None,
    'index': None,
    'IsOccluded': True,
    'IsTruncated': False,
    'IsGroupOf': False,
    'IsDepiction': False,
    'IsInside': False,
}>

Key annotation features:

• Bounding box coordinates (normalized)
• Object class labels
• Occlusion and truncation flags
• Group, depiction, and inside flags

Your Task

Your challenge is to curate a subset of this dataset that:

1. Reduces the overall size of the dataset
2. Maintains or improves the performance of an object detection model (YOLOv8m)

Remember, the goal is to find the optimal balance between dataset size and model performance, as measured by our evaluation metric:

Score = (mAP * log(N)) / N

Where mAP is the Mean Average Precision on our hidden test set, and N is the number of images in your curated dataset.

Additional Notes

• Ensure you comply with all dataset usage terms and conditions
• Do not use external data sources for this challenge

Good luck, and happy curating!

Citations

The Open Images v4 paper can be found here

@article{OpenImages,
  author = {Alina Kuznetsova and Hassan Rom and Neil Alldrin and Jasper Uijlings and Ivan Krasin and Jordi Pont-Tuset and Shahab Kamali and Stefan Popov and Matteo Malloci and Alexander Kolesnikov and Tom Duerig and Vittorio Ferrari},
  title = {Open Images Dataset V6: A Large-Scale Dataset for Object Detection in the Wild.}, {Open Images Dataset V7: A Large-Scale Dataset for Object Detection in the Wild.}
  year = {2020},
  journal = {IJCV}
}