InpaintCOCO - Fine-grained multimodal concept understanding (for color, size, and COCO objects)

Dataset Summary

A data sample contains 2 images and 2 corresponding captions that differ only in one object, the color of an object, or the size of an object.

Many multimodal tasks, such as Vision-Language Retrieval and Visual Question Answering, present results in terms of overall performance. Unfortunately, this approach overlooks more nuanced concepts, leaving us unaware of which specific concepts contribute to the success of current models and which are ignored. In response to this limitation, more recent benchmarks attempt to assess particular aspects of vision-language models. Some existing datasets focus on linguistic concepts utilizing one image paired with multiple captions; others adopt a visual or cross-modal perspective. In this study, we are particularly interested in fine-grained visual concept understanding, which we believe is not covered in existing benchmarks in sufficient isolation. Therefore, we create the InpaintCOCO dataset which consists of image pairs with minimum differences that lead to changes in the captions.

Download the dataset:

from datasets import load_dataset
dataset = load_dataset("phiyodr/inpaintCOCO")

Supported Tasks and Leaderboards

InpaintCOCO is a benchmark to understand fine-grained concepts in multimodal models (vision-language) similar to Winoground. To our knowledge InpaintCOCO is the first benchmark, which consists of image pairs with minimum differences, so that the visual representation can be analyzed in a more standardized setting.

Languages

All texts are in English.

Dataset Structure

DatasetDict({
    test: Dataset({
        features: ['concept', 'coco_caption', 'coco_image', 
            'inpaint_caption', 'inpaint_image', 
            'mask', 'worker', 'coco_details', 'inpaint_details', 'mask_details'],
        num_rows: 1260
    })
})

Data Instances

An example looks as follows:

{'concept': 'object',
 'coco_caption': 'A closeup of a large stop sign in the bushes.',
 'coco_image': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=512x512>,
 'inpaint_caption': 'A wooden bench in the bushes.',
 'inpaint_image': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=512x512>,
 'mask': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=512x512>,
 'worker': 'k',
 'coco_details': {'captions': ['A stop sign is shown among foliage and grass.',
     'A close up of a Stop sign near woods. ',
     'A closeup of a large stop sign in the bushes.',
     'A large oval Stop sign near some trees.',
     'a close up of a stop sign with trees in the background'],
   'coco_url': 'http://images.cocodataset.org/val2017/000000252332.jpg',
   'date_captured': '2013-11-17 08:29:48',
   'flickr_url': 'http://farm6.staticflickr.com/5261/5836914735_bef9249442_z.jpg',
   'height': 480,
   'id': 252332,
   'image_license': 'https://creativecommons.org/licenses/by/2.0/',
   'text_license': 'https://creativecommons.org/licenses/by/4.0/legalcode',
   'width': 640},
 'inpaint_details': {'duration': 18,
   'guidance_scale': 7.5,
   'num_inference_steps': 100,
   'prompt': 'wooden bench',
   'prompts_used': 2,
   'quality': 'very good'},
 'mask_details': {'height_factor': 25,
   'prompt': 'stop sign',
   'prompts_used': 1,
   'width_factor': 25}}

Dataset Creation

The challenge set was created by undergraduate student workers. They were provided with an interactive Python environment with which they interacted via various prompts and inputs.

The annotation proceeds as follows: The annotators are provided with an image and decide if the image is suitable for editing. If yes, they input the prompt for the object that should be replaced. Using the open vocabulary segmentation model CLIPSeg (Lüddecke and Ecker, 2022) we obtain a mask for our object of interest (i.e., "fire hydrant"). Then, the annotator inputs a prompt for Stable Diffusion v2 Inpainting (Rombach et al., 2022) (e.g. with the prompt "yellow fire hydrant"), which shows three candidate images. The annotators can try new prompts or skip the current image if the result is insufficient. Finally, the annotator enters a new caption that matches the edited image.

Source Data

InpaintCOCO is based on MS COCO 2017 validation set (image, annotations).

@misc{lin2015microsoft,
      title={Microsoft COCO: Common Objects in Context}, 
      author={Tsung-Yi Lin and Michael Maire and Serge Belongie and Lubomir Bourdev and Ross Girshick and James Hays and Pietro Perona and Deva Ramanan and C. Lawrence Zitnick and Piotr Dollár},
      year={2015},
      eprint={1405.0312},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Limitations

The images in the COCO dataset come from Flickr from 2014; therefore, they reflect the Flickr user structure at that time, i.e., the images mostly show the Western world and/or other countries from the Western perspective. The captions are in English. Thus, the model we developed does not generalize well beyond the Western world

Licensing Information

• Images come with individual licenses (image_license) based on their Flickr source. The possible licenses are
  • CC BY-NC-SA 2.0 Deed,
  • CC BY-NC 2.0 Deed,
  • CC BY 2.0 Deed, and
  • CC BY-SA 2.0 Deed.
• The remaining work comes with the CC BY 4.0 Legal Code license.

Citation Information

Our InpaintCOCO dataset:

@misc{roesch2024enhancing,
      title={Enhancing Conceptual Understanding in Multimodal Contrastive Learning through Hard Negative Samples}, 
      author={Philipp J. Rösch and Norbert Oswald and Michaela Geierhos and Jindřich Libovický},
      year={2024},
      eprint={2403.02875},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

For the MS COCO dataset please see above.