Skip to content

Benchmark guide

GazeForge uses multiple external benchmarks because no single dataset answers every validation question. This page helps you choose which evidence trail to inspect without ranking datasets or treating unlike sampling, annotation, split, and source conditions as interchangeable.

Check the canonical status before citing a benchmark

The current evidence class and scientific boundary for each benchmark are generated on Evidence status. This guide routes you to the right documentation; it does not promote evidence strength.

Choose a benchmark

  • Lund2013

    Start here for paired expert event labels, participant-held-out event modelling, and sampling/purity sensitivity. Lower-rate analyses are derived from native 500 Hz data and are not native GP3 validation.

    Lund2013 overview → · Sampling sensitivity →

  • Hollywood2EM

    Start here for expert-corrected event annotations and source-token-held-out evidence. Current source-token-disjoint evidence is not participant-disjoint, so participant-level generalisation must not be inferred.

    Hollywood2EM benchmark → · Source resolution →

  • Gaze-in-the-Wild

    Start here for reviewed participant-disjoint, task-agnostic event evidence and the associated source/task provenance trail. The complete authoritative numeric task mapping remains unresolved.

    Gaze-in-the-Wild benchmark → · Source resolution →

  • VISUS

    Start here for the current bounded partial public-derivative 60 Hz evidence and dynamic-AOI/source-recovery work. Do not treat it as full-dataset VISUS validation or native GP3 validity.

    VISUS public partial evidence → · Source resolution →

Choose by validation question

If your question is… Start with Why
How do event methods behave under controlled lower-rate derivation? Lund2013 expert labels plus explicit native/derived-rate separation
What does the source-token-held-out Hollywood2EM evidence support? Hollywood2EM preserves the non-participant-disjoint split boundary
Where is participant-disjoint task-agnostic event evidence available? Gaze-in-the-Wild participant-disjoint evidence with task-mapping limits kept visible
What empirical VISUS-derived material is publicly verified today? VISUS public partial evidence bounded public-derivative evidence without full-benchmark promotion
Does any benchmark establish native 60 Hz / GP3 validity? Native 60 Hz validation derived/external conditions must not substitute for native-device evidence
Which benchmark results are publication-gated? Frozen empirical evidence report fingerprints and publication gates are shown explicitly
Which source/rights questions remain unresolved? Source-resolution status central routing to benchmark-specific provenance work

Comparison rules

Before comparing numbers across benchmark pages:

  1. confirm the reference annotation is comparable;
  2. confirm the split unit (participant, source token, or another unit);
  3. separate native acquisition rate, processed-stream rate, and any derived analysis grid;
  4. check whether the result is Frozen, Reviewed, Bounded, pending, or another current status on Evidence status;
  5. preserve source/rights limitations rather than treating data availability as proof of authority.

A larger metric from a different benchmark design is not automatically stronger evidence. GazeForge therefore keeps benchmark-specific provenance beside the result instead of publishing a single cross-dataset league table.

Continue