Skip to content

VISUS public 60 Hz event-extension evidence

GazeForge contains a second, narrowly scoped empirical VISUS-derived evidence tranche built from two complete Tobii exports that are publicly present in the VISUS-supervised Maurice189/eye-slitscan repository.

This evidence is deliberately separate from the VISUS public partial 60 Hz AOI evidence and from the canonical VISUS Frozen Evidence gate. It expands real participant-level 60 Hz event coverage; it does not resolve the full VISUS benchmark.

Frozen source identity

The source repository is pinned to commit:

a8ea2402936122f9e5c98152460bd16a4ba97740

Two complete Tobii exports are treated as empirical data:

Participant Upstream path Bytes Git blob SHA-1 SHA-256
P5B core/importer/eye-tracker-output/test/Tobii_exports/01-OK.tsv 263,072 fd2371fd6f44de8a188e52439a0fea6b2054f975 df9ee65adb3a9872121f5ae3204842b0ab2efe107682151a287168f526b2c4b6
P3A core/importer/eye-tracker-output/test/Tobii_exports/02-OK.tsv 256,538 e39b0c0d2c50c22dec76b93581a4ca2bce784546 619da12969d04b774b9456b1b50e8cba6b21c04df33e2e90ef1c798a511d2bbc

The upstream C++ unit test is also pinned as provenance-only source material:

  • path: core/importer/eye-tracker-output/test/test.cc
  • Git blob SHA-1: d5f681eb4cc7b90c6078dc1fb7ceeccb4cc03c41
  • SHA-256: 999ab4c53e0817a65fc12bb9c143e19490059bb71bab5c6a03a2533cc9e5d1ee

That test explicitly loads 01-OK.tsv and 02-OK.tsv as valid Tobii exports and asserts known timestamps, fixation indices, fixation durations, and mapped coordinates. This establishes that the two files were intentional valid-input fixtures upstream rather than unreferenced loose files.

Three LibreOffice lockfiles in the same source directory are retained only as provenance, never as empirical observations. They identify historical files named P2B-03-dialog.tsv, P4B-03-dialog.tsv, and P6A-03-dialog.tsv, all opened by Maurice Koch on 21 May 2017. The underlying three TSV files are not present in the Git history recovered so far.

Empirical results

Both complete exports contain 1920×1080 media coordinates and sample intervals with a median positive microsecond delta of 16,625 µs, corresponding to approximately 60.1504 Hz.

Metric P5B P3A Aggregate
Samples 1,145 1,145 2,290
Both-eye-valid samples 1,136 1,136 2,272
Both-eye-valid fraction 0.9921 0.9921 0.9921
Inferred sampling rate 60.1504 Hz 60.1504 Hz 60.1504 Hz
Fixation events 62 43 105
Fixations with on-screen mapped point 62 42 104
On-screen fixation fraction 1.0000 0.9767 0.9905
Movie span 19.063 s 19.069 s —

The Tobii export settings recovered from both source headers are identical: Eye = Average, Validity = Normal, Fixation filter = Tobii fixation filter, Velocity threshold = 35, and Distance threshold = 35.

Fixation-duration interpretation

The exported fixation-duration values sum to 39,614 ms across both recordings, whereas the two MovieStart→MovieEnd spans sum to 38,132 ms. This is not interpreted as 39,614 ms of fixation time confined to the video segments. Tobii fixation durations can extend across segment boundaries; GazeForge therefore freezes these values as exported event metadata and explicitly records that they are not clipped to the movie boundaries.

03-dialog is a strong inference, not resolved identity

The two source files are named 01-OK.tsv and 02-OK.tsv; they do not themselves contain a VISUS stimulus filename. Consequently GazeForge does not promote the stimulus identity to resolved status.

The candidate 03-dialog assignment is nevertheless strongly supported by converging evidence:

  1. both complete movie segments are approximately 19 s long (19.063 s and 19.069 s);
  2. the original VISUS benchmark describes the Dialog stimulus as 19 s;
  3. the same upstream directory contains three historical lockfiles explicitly named for 03-dialog recordings;
  4. Osnabrück University's WACV 2017 converted-dataset page independently lists Kurzhals stimulus K3 as Dialog.

The immutable evidence record therefore uses:

identity_status = strongly-inferred-not-file-bound

and requires:

stimulus_identity_resolved = false

No Dialog AOI annotation stream has been recovered for these two exports, so this tranche contains event/QC evidence rather than dynamic-AOI performance evidence.

Analysis-use boundary

The current Osnabrück University WACV 2017 dataset page states that the converted datasets are provided for research purposes, asks users to cite the conversion work and original dataset, and lists the Kurzhals dataset as an 11-video converted collection:

https://www.ikw.uni-osnabrueck.de/en/research_groups/computer_vision/research/interactive_3d_modelling/multimedia_container/wacv17.html

This supports an analysis-use basis. It does not by itself establish a dataset-specific unrestricted redistribution license for the original VISUS bytes. GazeForge therefore stores hashes, provenance, and derived metrics but does not vendor the upstream TSV files.

Reproducibility

The live workflow runs:

python scripts/visus_public_event_extension_probe.py

It re-downloads the exact pinned files, independently reconstructs each Git blob SHA-1, checks byte sizes, computes SHA-256 identities, parses the Tobii exports, validates the upstream unit-test assertions and lockfile provenance, recomputes participant/event metrics, and produces a deterministic JSON probe.

Frozen probe identity:

  • workflow run: 33928088952
  • probe head: 9308c79dc3f6ef9d42383e85ec2abc6bad0b783d
  • probe fingerprint: 47316bcb77e3cf1a92fdb95df84cf401c31fdafd6fc6affce3b9f6405f92312e
  • artifact: 9957528541
  • artifact ZIP SHA-256: cc9f139c9c7b27649fefd8a302801707bbe6dc6140f49fafd3525971a02f437d

Immutable evidence fingerprint:

2f12bd83d71786bfae7101dec6515c49c5ff4e696df8675b3955300e5e5e6dfd

The committed validator also binds a newly generated live probe back to this frozen evidence contract. Any drift in source identities, participant metrics, provenance, or scientific boundaries fails closed.

What this evidence does not establish

This tranche does not establish any of the following:

  • exact file-bound 03-dialog stimulus identity for P5B or P3A;
  • Dialog AOI annotations or AOI-hit metrics;
  • the complete original 25-participant × 11-stimulus VISUS corpus;
  • a resolved unrestricted redistribution license;
  • independent human-human annotation agreement;
  • model-human validation;
  • native Gazepoint GP3 evidence;
  • canonical VISUS Frozen Evidence.

Those requirements remain open and are intentionally protected by the validator.