Short description (use in GitHub About): Interactive companion site for the DeepBench paper: domain-specific robustness evaluation of Vision–Language Models (VLMs) with live corruption galleries, per-domain experiments (RQ1–RQ6), and copyable citations.
Live demo
- GitHub Pages: https://ml-lab-htw.github.io/deepbench_paper/
Overview This repository contains a static, self-contained web site (HTML/CSS/JS) that presents the DeepBench paper’s results in an interactive format:
- Corruption gallery with thumbnails and 4-up severity examples
- Use Cases panel showing datasets and example images
- Experiments panels RQ1–RQ6 with plots per domain and optional per-corruption filtering
- Expandable Sources parsed from a BibTeX file (paper-style, no links)
- “How to Cite” section with a light-themed, directly copyable BibTeX block
- Graphical abstract and highlights
- Modern sticky UI: shrinking page title, compact nav, sticky corruption thumbnails
Key features
- Corruption Examples
- Click thumbnails to switch corruption; shows four examples (severity levels) from assets/images//
- Supports a dynamic manifest at assets/images/manifest.json; falls back to conventional filenames
- Use Cases
- Dedicated panel with thumbnails, short domain descriptions, and 4 example images per use case
- Experiments (RQ1–RQ6)
- RQ1: LLM-based corruption selection
- RQ2: Main robustness experiment across models and domains
- RQ3: Domain specialists
- RQ4: Architecture ablations
- RQ5: Pretraining data effects (Acc + MCE/Flip)
- RQ6: Label-Flip Probability correlation with accuracy
- Per-domain plots live under assets/experiments//
- Sources and Citations
- BibTeX loaded from assets/references.bib and rendered as paper-style references (expandable)
- Copyable BibTeX in “How to Cite,” light code theme with a one-click copy button
- Polished UX
- Sticky shrinking title + compact menu
- Sticky corruption thumbnails that span multiple experiment sections
- Smooth scrolling and automatic anchor offset so section headers aren’t covered
- “Coming Soon” overlay toggle: ?comingSoon=1 or window.setComingSoon(true)
Project structure
- index.html – Main page
- css/style.css – Styles (header/menu, sticky title, thumbnails, code block theme, etc.)
- js/main.js – Interactive logic (thumbnails, plots, sticky behavior, BibTeX rendering)
- assets/
- references.bib – Paper references (BibTeX)
- images/ – Corruption examples and thumbnails (supports manifest.json)
- experiments/ – Plots for RQ1–RQ6 (grouped by experiment)
- CVIU_*.pdf, LaTeX sources, tables, and section .tex files used for content
Local preview You can serve the site locally with any static server. For example (Python 3):
cd deepbench_paper
python3 -m http.server 8000
# Open http://localhost:8000Publishing with GitHub Pages (via Actions) This repository is pre-configured to deploy using GitHub Pages + Actions.
- Enable Pages for this repo
- On GitHub: Settings → Pages → Build and deployment → Source: GitHub Actions
- Push to main
- Every push to the main branch triggers the workflow at .github/workflows/deploy.yml
- The deployed URL appears under Settings → Pages and in the “Deploy to GitHub Pages” step logs
Manual gh-pages alternative (optional) If you prefer a branch deployment:
- Create a gh-pages branch with this site at the repo root
- Settings → Pages → Deploy from a branch → gh-pages / (root)
Editing content
- Add/replace plot images under assets/experiments//
- Add or adjust corruption example images under assets/images//
- Optional: list the exact example filenames in assets/images/manifest.json to override auto-fallbacks
- Update the bibliography in assets/references.bib
- Update textual content in index.html (graphical abstract, highlights, RQ summaries) and css/style.css/js/main.js as needed
Useful toggles and behaviors
- Coming Soon overlay
- Enable via URL: ?comingSoon=1 or programmatically: window.setComingSoon(true)
- Sticky thumbnails bar
- Automatically sticks from “Corruption Examples” through the end of RQ6
- Thumbnails shrink by 50% while sticky
- Anchor offset
- Sections use scroll-margin-top driven by a dynamic header offset so titles aren’t hidden under the sticky header
Suggested repository topics (GitHub → About → Topics)
- computer-vision, robustness, benchmark, corruptions, vision-language-models, clip, siglip, llava, gemma, laion, datacomp, label-flip-probability, deepbench
Issues and contributions
- Please open issues for bugs or enhancement requests
- Pull requests welcome for plots, content, styles, or small JS improvements