Findable
Names the public dataset, exact source rows, access date and deterministic seed.
INFOSCI 301
Preliminary redesign
13 Sep 2026
What can a game
look like?
A visual discovery interface for DKU students beginning to explore technical art.
Instead of treating genre as visual style, this prototype places gameplay frames by computed visual similarity. Select any frame to inspect its visual neighborhood.
Loading the governed 40-image sample…
01
Discovery, not classification. The map offers examples to inspect; it does not claim to define a universal vocabulary of game styles.
Interactive figure 01
Screenshot Style Map
Nearby frames share low-level colour, texture and composition features. PCA axes have no standalone semantic meaning.
Computed attributes
The visual sample could not be loaded.
Inspect the dataset and method links below, or reload the page.
How to read this figure
Supported: frames can be compared within this deliberately small sample using documented image features.
Not supported: proximity does not prove that two games share an art style, production method, or required technical-art skill.
Field evidence · 04 Sep 2026
From overview to inspectable example
Two Shanghai visits suggested that abstract information becomes more legible when it remains connected to a concrete object, scene or action.
A digital overview and projected traffic information remained physically connected to a site model.
Physical controls connected small inspectable scenes with a larger visual display.
02
Design consequence. The redesign connects the overview map to concrete frames: every point is an image a learner can select, inspect and compare.
Data, evidence & governance
Know what the map knows.
Transparency is part of the interface, not supplementary decoration.
Accessible
Ships a small local sample and links back to the public source rather than loading 2.69 GB at runtime.
Interoperable
Exports image IDs, source rows, features, coordinates, clusters and neighbours as JSON.
Reusable
Documents sampling and feature generation in a rerunnable script, alongside reuse caveats.
Source & transformation
What was used
Bingsu/Gameplay_Images contains 10,000 gameplay frames across ten game-identity classes. This prototype uses four deterministic rows per class (seed 301), resized for the web.
The dataset card reports CC-BY-4.0. It also says frames were saved from YouTube gameplay videos; per-frame video URLs and rights-holder verification are not provided. The prototype therefore preserves attribution and does not treat the card as proof of all underlying image rights.
Fairness & uncertainty
What remains unknown
Ten popular games from a 2022 upload do not represent the wider game ecosystem, regional production, independent games, or DKU learner priorities. Equal sampling balances this interface sample, not the world.
Descriptors such as “dark” and “vivid” are thresholded image measurements. Cluster numbers are computed groupings. Neither is a verified art-style annotation.
CARE applicability
Authority without overclaiming
No Indigenous or collectively governed community data is knowingly included, so the project does not claim CARE compliance. CARE’s emphasis on authority informs a practical boundary: public labels, computed interpretations and future DKU student feedback remain visibly separate.
Redesigned IEEE VIS paper
ImageSI: Semantic Interaction for Deep Learning Image Projections
This prototype adapts ImageSI’s interaction principle—using selected images to guide visual discovery—to a small learning interface. It does not reproduce or claim the paper’s deep-learning method.
Official IEEE VIS 2024 page ↗