Table of Contents in HTML: How Mid-2000s Researchers Structured Digital Proceedings

The 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2005) published its proceedings as a set of HTML pages, each paper accessible via a master index page that listed every title, author, and session. That master index was nothing more than a table of contents (TOC) built from nested unordered lists and hyperlinked anchors. For the pattern recognition community, this simple HTML structure became the default navigation interface for digital libraries, replacing printed bookmarks and PDF outlines. Understanding how that TOC was generated—and the markup decisions behind it—reveals a forgotten piece of mid-2000s information engineering.

The Anatomy of a Conference TOC Page

A typical HTML TOC for a 2005 proceedings looked like this:

<ul>
  <li><a href="paper001.html">Session 1: Object Detection</a>
    <ul>
      <li><a href="paper001.html#intro">Introduction</a></li>
      <li><a href="paper001.html#method">Method</a></li>
    </ul>
  </li>
</ul>

Each paper’s own page used <a name="..."> or <h2 id="..."> to create fragment identifiers. The TOC linked to those fragments, allowing readers to jump directly to the method section or the experimental results. This practice predated widespread use of JavaScript-based scrolling and was entirely server-generated, often from LaTeX source files that were converted to HTML via tools like latex2html or TeX4ht.

Why Mid-2000s Researchers Cared About HTML TOCs

Three factors drove the adoption of HTML-based tables of contents in academic pattern recognition circles:

• Bandwidth constraints: Downloading a full PDF (often 2–5 MB) was slow on university networks. An HTML TOC with inline links let researchers preview abstracts and figures without fetching the entire document.
• Cross-referencing: Conference proceedings aggregated dozens of papers. A well-structured TOC enabled quick scanning of sessions—e.g., “Stereo Reconstruction” or “Face Recognition”—and direct access to relevant papers.
• Accessibility: Screen readers and text-based browsers (still used in 2005 by some visually impaired researchers) could parse an HTML list far more reliably than a scanned PDF.

One notable example was the International Conference on Acoustics, Speech, and Signal Processing (ICASSP) 2006 proceedings, which used a two-level TOC: top-level sessions (e.g., “Speech Recognition”) and second-level paper titles. The HTML was hand-tuned by editorial staff to ensure consistent anchor naming across all 1,200+ papers. This manual effort highlighted a recurring problem: generating a reliable TOC from heterogeneous source documents.

The Generation Pipeline: From LaTeX to HTML

Most pattern recognition conferences in the mid-2000s accepted papers in LaTeX. The conversion pipeline for the proceedings TOC typically followed these steps:

1. Authors submitted a LaTeX source with section{...} and subsection{...} commands.
2. The proceedings editor ran a script that extracted section headings and their page numbers from the compiled PDF (using pdftotext or pdfinfo).
3. An XSLT transformation converted the extracted data into an HTML unordered list, assigning each heading a unique id based on the paper number and section number (e.g., p123sec2).
4. The HTML TOC was inserted into a frameset or a sidebar iframes were used—a technique that later fell out of favor due to accessibility issues.

This pipeline was fragile. If an author used custom LaTeX macros or non-standard sectioning commands, the script would miss headings. Some proceedings resorted to manual XML authoring, where editors typed the TOC entries by hand. The 2004 European Conference on Computer Vision (ECCV) proceedings, for instance, included a handwritten HTML TOC that occasionally had broken links—a problem that the community tolerated as a minor nuisance.

Semantic Markup vs. Visual Presentation

Mid-2000s HTML TOCs were rarely styled with CSS. The default browser rendering—bulleted list with blue underlined links—was considered acceptable. But some organizers experimented with richer markup. The 2005 International Conference on Machine Learning (ICML) used a <table> layout for its TOC, with columns for paper ID, title, and author. While visually compact, this approach lost the semantic structure that a <ul> provided. Screen readers would announce each cell without context, making navigation harder.

A more forward-looking approach appeared in the 2006 Workshop on Applications of Computer Vision (WACV): the TOC was built using definition lists (<dl>), with <dt> for the paper title and <dd> for the author list. This preserved the relationship between title and metadata while remaining machine-readable. However, definition lists were rarely used elsewhere, so the pattern did not spread.

The Role of Anchor Naming Conventions

Consistent anchor naming was essential for cross-linking between papers. A typical convention was:

• p{paperNumber}sec{sectionNumber} for major sections (e.g., p042sec1).
• p{paperNumber}fig{figureNumber} for figures (e.g., p042fig3).
• p{paperNumber}eq{equationNumber} for equations (e.g., p042eq7).

This naming scheme allowed the TOC to link not only to papers but also to specific figures and equations. In practice, few readers used these deep links because the TOC only listed sections, but authors occasionally referenced another paper’s figure directly in their own HTML version. The practice was documented in the “Author Guidelines for Electronic Proceedings” published by the IEEE Computer Society in 2004, which recommended anchor names of the form sec:paperID:sectionNumber.

Impact on Object Tracking and Speech Research

The HTML TOC structure directly influenced how researchers consumed new work. For example, a researcher working on high-speed video object tracking could open the CVPR 2005 TOC, scan the “Motion and Tracking” session, and click directly to a paper’s “Experimental Setup” section. That kind of targeted navigation was faster than scrolling through a PDF. In fact, the efficiency of HTML TOCs contributed to the rapid dissemination of results in fields like object tracking—a topic we explored in our earlier post on 120 Frames per Second: How High-Speed Video Reshaped Object Tracking in the Mid-2000s.

Similarly, speech recognition researchers relied on HTML TOCs to jump to acoustic feature descriptions. The Zero Crossing Analysis: The Forgotten Acoustic Feature of Mid-2000s Speech Recognition article notes how ZCR papers were often buried in long proceedings; a well-structured TOC made them discoverable.

The Decline of Handcrafted HTML TOCs

By 2008, digital libraries such as IEEE Xplore and ACM DL had taken over the role of conference proceedings distribution. These platforms generated their own dynamic TOCs from metadata, rendering the static HTML TOC pages obsolete. The server-generated <ul> of the mid-2000s was replaced by JavaScript-driven menus and search filters. However, the underlying principle—using fragment identifiers to create deep links into papers—persists in modern PDF viewers and HTML5 document outlines.

One concrete legacy is the id attribute convention still used in arXiv HTML abstracts. Every arXiv paper page includes an #abstract anchor, and many authors manually add ids to sections. The mid-2000s practice of generating those anchors from LaTeX section numbers directly influenced the arXiv submission guidelines. Today, a researcher can link to and jump to the second section—a direct descendant of the CVPR 2005 TOC pipeline.