metadata

task_categories:
  - object-detection
dataset_info:
  features:
    - name: image_id
      dtype: string
    - name: image
      dtype: image
    - name: label_bbox
      list:
        - name: bbox
          sequence: int64
        - name: bbox_id
          dtype: string
        - name: label
          dtype: string
    - name: issues
      list:
        - name: confidence
          dtype: float64
        - name: description
          dtype: string
        - name: issue_type
          dtype: string
  splits:
    - name: train
      num_bytes: 13436697177
      num_examples: 82081
    - name: validation
      num_bytes: 6606403140
      num_examples: 40137
    - name: test
      num_bytes: 6653024122
      num_examples: 40775
  download_size: 26617129269
  dataset_size: 26696124439
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*

COCO-2014-VL-Enriched

An enriched version of the COCO 2014 dataset with label issues! The label issues helps to curate a cleaner and leaner dataset.

Description

The dataset consists of 6 columns:

image_id: The original image filename from the COCO dataset.
image: Image data in the form of PIL Image.
label_bbox: Bounding box annotations from the COCO dataset. Consists of bounding box coordinates, confidence scores, and labels for the bounding box generated using object detection models.
issues: Quality issues found such as duplicate, mislabeled, dark, blurry, bright, and outlier images.

Usage

This dataset can be used with the Hugging Face Datasets library.:

import datasets
ds = datasets.load_dataset("visual-layer/coco-2014-vl-enriched")

Interactive Visualization

Visual Layer provides a platform to interactively visualize a dataset and highlight quality issues such as duplicates, mislabels, outliers, etc. Check it out here. No sign-up required.

License & Disclaimer

We provide no warranty on the dataset, and the user takes full responsibility for the usage of the dataset. By using the dataset, you agree to the terms of the ImageNet-1K dataset license.

About Visual Layer