README.md · gokaygokay/OPA

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: mask
      dtype: image
    - name: fg_id
      dtype: string
    - name: bg_id
      dtype: string
    - name: position
      dtype: string
    - name: scale
      dtype: float32
    - name: label
      dtype: int32
    - name: image_filename
      dtype: string
    - name: mask_filename
      dtype: string
    - name: background
      dtype: image
    - name: foreground
      dtype: image
    - name: category
      dtype: string
  splits:
    - name: train
      num_bytes: 15378173089.116
      num_examples: 62074
    - name: test
      num_bytes: 2802618925.656
      num_examples: 11396
  download_size: 15201692849
  dataset_size: 18180792014.772
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
license: mit
language:
  - en
pretty_name: OPA Composite
size_categories:
  - 10K<n<100K

Dataset Card for new_OPA_dataset

Dataset Summary

The Object-Placement-Assessment (OPA) dataset is a synthesized dataset designed for evaluating the rationality of object placement in composite images. It is based on the COCO dataset, featuring foreground objects pasted onto compatible background images with varying sizes and locations. Each composite image is annotated with a binary label (0 or 1) indicating whether the object placement is reasonable, considering factors like location, size, occlusion, semantics, and perspective.

This dataset contains 62,074 training images and 11,396 test images, with no overlap in foregrounds or backgrounds between the splits. It is ideal for tasks such as object placement prediction, image composition, and visual common sense reasoning.

Supported Tasks and Leaderboards

Task: Object Placement Assessment
Description: Predict whether a composite image’s object placement is rational (label 1) or irrational (label 0) based on visual and semantic cues.
Leaderboard: No public leaderboard is currently available.

Languages

The dataset includes image data with English metadata (e.g., category names, file names).

Dataset Structure

Data Instances

Each instance represents a composite image with its associated mask and metadata. An example from the train split looks like:

{
    'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=640x480>,
    'mask': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=640x480>,
    'fg_id': '4535',
    'bg_id': '367260',
    'position': '[482, 281, 151, 126]',
    'scale': 0.26269999146461487,
    'label': 0,
    'image_filename': '4535_367260_482_281_151_126_0.2627_0.jpg',
    'mask_filename': 'mask_4535_367260_482_281_151_126_0.2627_0.jpg'
}

Data Fields

image: The composite image (JPEG, typically 640x480 pixels).
mask: The mask of the foreground object in the composite image (JPEG, same size as the image).
fg_id: Identifier for the foreground object (string).
bg_id: Identifier for the background image (string).
position: Bounding box of the foreground object in the format [x, y, w, h], where x, y is the upper-left corner, and w, h are width and height (string).
scale: The maximum ratio of foreground width/height to background width/height (float32).
label: Binary label indicating placement rationality (0 for irrational, 1 for rational, int32).
image_filename: File name of the composite image (string).
mask_filename: File name of the mask image (string).

Data Splits

Train: 62,074 images (21,376 positive, 40,698 negative samples).
Test: 11,396 images (3,588 positive, 7,808 negative samples).

There is no overlap in foregrounds (2,701 unique in train, 1,436 in test) or backgrounds (1,236 unique in train, 153 in test) between splits.

Dataset Creation

Curation Rationale

The OPA dataset was created to support research in object placement assessment, addressing challenges in image composition by evaluating the plausibility of foreground object placements. It was synthesized from COCO by selecting unoccluded objects, pasting them onto compatible backgrounds, and annotating rationality via human labeling.

Source Data

Initial Data Collection: Derived from the COCO dataset, with foreground objects and backgrounds curated for compatibility.
Annotation Process: Composite images were generated with random sizes and locations, then labeled by human annotators for rationality.

Annotations

Annotation Types: Binary rationality labels (0 or 1), bounding box positions, and scale values.
Process: Human annotators assessed each composite image for placement plausibility based on location, size, occlusion, semantics, and perspective.

Considerations for Using the Data

Social Impact of Dataset

This dataset advances research in image composition and visual reasoning, with applications in augmented reality, content creation, and automated design. However, biases in the COCO dataset (e.g., underrepresentation of certain scenes or objects) may carry over.

Discussion of Biases

As a derivative of COCO, the dataset may inherit biases related to object categories, scene diversity, or cultural contexts. Users should evaluate its suitability for specific applications.

Other Known Limitations

The dataset focuses on binary rationality labels, which may not capture nuanced placement quality.
Negative samples include specific issues (e.g., inappropriate size, occlusion), but the dataset does not categorize these issues explicitly.

Additional Information

Dataset Curators

The OPA dataset was created by Liu Liu, Zhenchen Liu, Bo Zhang, Jiangtong Li, Li Niu, Qingyang Liu, and Liqing Zhang from the Brain Cognition & Machine Intelligence Lab (BCMI).

Licensing Information

The dataset is released for research purposes. As a derivative of COCO, it inherits COCO’s licensing terms (CC BY 4.0). Users should verify compliance with COCO’s license.

Citation Information

If you use this dataset, please cite the original OPA paper:

@article{liu2021OPA,
  title={OPA: Object Placement Assessment Dataset},
  author={Liu, Liu and Liu, Zhenchen and Zhang, Bo and Li, Jiangtong and Niu, Li and Liu, Qingyang and Zhang, Liqing},
  journal={arXiv preprint arXiv:2107.01889},
  year={2021}
}