Brainhack School Project: Image10K

My name is Sara, and I'm currently completing my undergraduate degree in Cognitive Neuroscience. I will begin a Master’s program in Psychology at the University of Montreal, where I’ll be conducting research in the Montembeault Laboratory.

As part of Brainhack School, I aim to create an interactive web-based showcase for the Images10k dataset — a large collection of behaviorally validated, naturalistic images spanning over 15 semantic categories. The goal of my project was to explore how we can make rich, complex datasets more accessible and meaningful to both researchers and the public.

Through the site, I present the dataset in an interactive and structured format:

• Visitors can browse the image collection using carousel galleries
• Explore scrollable previews of image-level metadata
• View interactive visualizations and analyses drawn from the dataset

This project allowed me to deepen my skills in data organization, visualization, and reproducible workflows, using tools like Jupyter Book, MyST Markdown, and Python. More broadly, it helped me engage with a growing challenge in cognitive neuroscience and AI — how to bridge ecological validity and computational modeling through large-scale, behaviorally grounded datasets.

Project Background & Goal

Many studies in cognitive neuroscience still use really simplified or artificial images, like objects on plain backgrounds.That can limit ecological validity and create a disconnect between what you see in the lab and real world perception. In contrast to that, Images10K provides naturalistic scenes, showing objects in their natural environments, which makes it easier to study how people understand visual scenes in real-world contexts. This approach fits with recent work that emphasizes the importance of using more realistic images in research (Hosu, Lin, Szirányi, & Saupe, 2019).

This project is based on the Images10K dataset — containing over 8,000 naturalistic images, annotated by human participants on the Zooniverse platform. The overall goal is to provide a well-organized and richly annotated image set (8,382 images, 15 semantic categories) that can be reused for training visual recognition models in AI and neuroscience.

My Objective

My personal contribution to this project focuses on creating a clear, structured, and interactive way to explore the Images10K dataset. Specifically, I worked on:

• Organizing the dataset:
  I used Datalad to retrieve and manage the dataset in a reproducible way. I then cleaned and structured the image-level metadata to ensure consistency and ease of access. This involved preparing files for web-based display, grouping images into their semantic categories, and excluding any redundant or problematic entries to keep the dataset clean and reusable.

• Visualizing the images and their associated metadata:
  I developed interactive components that allow users to explore the dataset visually. This includes scrollable metadata tables, carousels, filtered views, and category-based displays that make the data more intuitive and engaging.

• Building an interactive site to showcase the categories and image properties:
  Using tools like Jupyter Book, MyST Markdown, and Python, I created a static web-based showcase that integrates image carousels, scrollable data previews, and exploratory visualizations. The goal was to present the dataset in a way that is both accessible and informative for researchers, students, and collaborators from different disciplines.

Tools I used

• GitHub for version control and to organize the project in a clear, collaborative, and shareable format.
• DataLad to retrieve and manage the dataset in a reproducible way.
• Python scripts to filter images by category, convert metadata formats, and prepare image paths for display.
• Jupyter Notebooks to explore the metadata, test visualizations, and generate previews of the dataset.
• Dash Bootstrap Components to build interactive UI elements like carousels and dropdown menus for browsing images.
• MyST Markdown to structure the content of the interactive website and document the project cleanly within Jupyter Book.
• Ubuntu terminal to navigate the file system, run scripts, and better understand how to work with files and folders at the command line.

The Data

The dataset includes:

• Over 8,000 naturalistic images, stored in semantically organized folders based on object categories
• Annotated labels for each image, including category, subcategory, and file path
• High-level semantic classifications (e.g., living vs. non-living, natural vs. artificial)
• Rich image-level metadata such as:
  • Number of participant annotations (via Zooniverse)
  • Inter-participant agreement scores
  • License information and usage permissions
  • Source URLs and author attributions

Concepts Implemented

• Interactive image carousels:
  Built using Dash and Dash Bootstrap Components to allow users to browse images by semantic category. Each carousel presents multiple images grouped by type (e.g., “birds,” “tools”), providing a visual overview of the dataset.

• Semantic filtering:
  Implemented category-based navigation using dropdown menus and structured folders, enabling users to explore subsets of the dataset (e.g., “animal,” “object,” “living,” “non-living”).

• Interactive visualizations:
  I used pandas and IPython.display to generate visual summaries of the metadata, including category distributions, agreement scores, and image sources.

• Scrollable metadata tables:
  Created separate pages with scrollable tables for each semantic category. These tables include information such as image ID, label, source URL, license type, and behavioral annotation statistics.

• Integrated statistical plots:
  Incorporated pre-existing plots generated using matplotlib and seaborn by other contributors working with the Images10K dataset. These include scatter plots and histograms that illustrate participant responses and highlight key behavioral and semantic patterns in the data.

• MyST Markdown and Jupyter Book:
  Used to structure the site and documentation, combining notebooks and Markdown files into a clean, open-access, and interactive website.

Delivery

• A structured GitHub repository with scripts, metadata, and documentation
• Jupyter Notebooks for metadata preview and exploratory analysis
• An interactive web interface built with Jupyter Book and MyST Markdown
• Image carousels and scrollable metadata tables for category-based exploration
• Downloadable metadata files hosted via Google Drive

Coming Soon: Public Web Version

Right now, the project runs locally on my computer. I am currently working on making the full webpage accessible online, so that others can explore the dataset and carousels without needing to run any code.

Link to view google slide presentation

presentation

Link to the Image10k-Compendium Website

You can view the full project and explore the carousels here:
Image10k Compendium Website

Note Only a limited set of images is included on the website due to GitHub storage and bandwidth constraints.

Note The interactive carousels (Dash apps) require Python to run, so they won’t launch directly in the website view.

To view the carousels interactively on the Website:

1. Clone this repository:
   Images10k-compendium

2. Install dependencies:
   pip install -r binder/requirements.txt

3. Run both notebooks in the content/ folder:

   • animated_being.ipynb
   • Objects.ipynb

References

Hosu, V., Lin, H., Szirányi, T., & Saupe, D. (2019). KonIQ-10k: An ecologically valid database for deep learning of blind image quality assessment. arXiv preprint arXiv:1910.06180. https://arxiv.org/abs/1910.06180