Skip to content

Gaze-in-the-Wild exact ProcessData structure

GazeForge has a reviewed structural/provenance record for the 68 exact original-publication ProcessData files in the official Gaze-in-the-Wild Figshare deposit. This gate is narrower than model validation: it establishes exact-file structure, processed timestamp-grid rates, recording-token pairing with distributed LabelData, and the applicability of already-reviewed first-party point-of-regard semantics to the exact distributed ProcessData structure.

Exact source scope

The reviewed source is Figshare project 74580:

  • ProcessData: article 11673645, DOI 10.6084/m9.figshare.11673645.v1;
  • LabelData: article 11673696, DOI 10.6084/m9.figshare.11673696.v1;
  • ProcessData_cleaned: article 11673717, deliberately excluded because it is not the original-publication distribution.

The 68 ProcessData files are bound to the previously reviewed exact-byte evidence and a dedicated SHA-256 ledger. Raw MAT files are streamed one at a time, verified, inspected, and deleted immediately; the repository and workflow artifact contain only JSON evidence/summary material.

Structural result

All 68 exact original ProcessData files passed the reviewed gate:

  • frozen Figshare size and MD5 match;
  • previously reviewed SHA-256 identity matches;
  • filename PrIdx / TrIdx agrees with the internal MATLAB fields;
  • ProcessData.T is strictly increasing;
  • ETG.POR is structurally compatible with the adapter and has N×2 shape;
  • ETG.Confidence and structural ETG.Labels match the timestamp count;
  • ETG.SceneResolution is 1920×1080 in every file;
  • no top-level LabelData object is embedded in a ProcessData file.

The exact distribution contains 68 recording tokens across 20 participant indices.

Processed timestamp-grid rate

Every exact file stores ProcessData.SR = 300. Independently, the median inverse delta of the strictly increasing ProcessData.T grid yields:

Quantity Hz
Minimum 299.98859845584474
Median 299.9948438624359
Maximum 299.9976557817862

These values describe the processed ProcessData.T timestamp grid. They are not promoted to the eye-tracker acquisition-hardware cadence. GazeForge continues to keep published acquisition provenance and processed-stream timing provenance separate.

The complete 68-file processed-rate ledger is frozen in:

validation/evidence/gaze-in-wild/gaze-in-wild-processdata-processed-rate-ledger-v1.json

LabelData ↔ ProcessData pairing

The 50 distributed LabelData files represent 37 unique PrIdx/TrIdx recording tokens across 16 participants and labeller indices 1, 2, 3, 5, 6.

Every one of those 37 labelled recording tokens has an exact matching ProcessData recording token. The distribution also contains 31 ProcessData recording tokens without separately distributed LabelData.

This establishes that task-agnostic participant-disjoint input partitioning is structurally available for the labelled subset. It does not create a participant-disjoint model-validation result.

Point-of-regard semantics

The structural result is bound to the previously reviewed first-party POR provenance evidence. At pinned first-party commit 52262d44e366a53369e10ca73c5f41daf0e8f1e5, ReadData_function.m shows that ETG.POR derives from Pupil norm_pos_x / norm_pos_y, with the y coordinate transformed as 1-y for MATLAB image convention, while scene resolution is retained separately.

Because all 68 exact original ProcessData files contain the reviewed POR/resolution structure, the first-party normalized scene-image POR semantics can be bound to the exact distribution. GazeForge's canonical conversion remains explicit:

  • x_px = por_x × scene_width_px
  • y_px = por_y × scene_height_px

This does not claim that corpus-wide observed POR value ranges have been empirically audited. That remains a separate boundary.

What this does not establish

This evidence does not establish or authorize:

  • acquisition-hardware cadence from the processed ~300 Hz grid;
  • a universal TrIdx → publication task mapping;
  • exact file-to-publication-task assignment;
  • task-stratified model validation;
  • participant-disjoint GazeForge model validation;
  • cross-dataset validation;
  • GP3 or native 60 Hz validity;
  • corpus-wide observed POR-range validation;
  • quarantine exit;
  • any new model-performance claim.

The publication-level participant/task matrix is separately reviewed, but TrIdx remains a distribution-native acquisition/processing token rather than a universal task name. No task labels are inferred here.

Reproducibility identities

Reviewed evidence:

  • exact ProcessData structure evidence: cffcc8a10d176cc5eb8c15c080cf450e3e1b1404ced00ea8907cdd4ee3c3517f;
  • exact ProcessData SHA-256 ledger: 85f131389315a1185e3a8973629c8e5ee3417bb73704de8054366e508af0a2e2;
  • ProcessData SHA-256 manifest: 162bf688fbd0bfdaf79a11d423f3689abd31bea9a4e8423bc03020ebd598354d;
  • processed-rate ledger: 1bd6643bd0c75faec9d070f74fb719f9178caa2da7e90bcbe356a0dbc5f38905;
  • stable per-file structure manifest: d82c3db448d71676240dabc66452fa7ef3c030f8390e3b0bb0fa56b30392e7fb;
  • stable reviewed corpus-structure fingerprint: 86c4c94742ada7eceee3d44ef9c1c11a81621acd533bbb4e28aef3f6e3976443;
  • parent exact-byte evidence: dbf277d698266fe53e337a15fc342e9af1835460d2bcc69c3992635de8fa0a00;
  • parent POR semantics evidence: 5eb245979385283571d6d9a5dda9a5d74dea6c00165bbe3a2587fcfa80e9adfb.

Discovery run 34199996646 / job 101976365230 produced artifact 10045524274, ZIP SHA-256 e51138b34c6e7c4ddcc8fa3ffd825884485019bf05270072d6abee97990f638e, and discovery probe fingerprint bbd7b0498f4d3b48e435bb0a79efd72d794e59e38fa31ccc73c2577bb95dd66b.

The dedicated reproduction workflow additionally validates a fresh 68-file probe against the stable per-file structure manifest and the frozen 68-row rate ledger, so retry-count transport noise cannot substitute for scientific identity.