Gaze-in-the-Wild task-mapping corroboration¶
GazeForge keeps publication task identity, first-party extraction structure, secondary numeric corroboration, and authoritative distributed file-to-task mapping as separate evidence layers. The reviewed exact Gaze-in-the-Wild participant-disjoint benchmark therefore remains task-agnostic until a first-party or otherwise authoritative source explicitly verifies the distributed TrIdx → publication-task mapping.
Current evidence ladder¶
| Evidence question | Current status |
|---|---|
| Are the four publication tasks known? | Yes — indoor navigation, ball catching, visual search, tea making |
| Does first-party RIT code preserve task-separated GIW extraction directories? | Yes |
Does that first-party code read PrIdx and TrIdx from each task directory's ProcessData and use them to address raw recordings? |
Yes |
Are TrIdx 1–3 numerically corroborated by an independent downstream study? |
Yes — ACE-DNV |
Does a reviewed first-party source explicitly publish the complete numeric TrIdx → task lookup? |
No |
Is TrIdx 4 → Tea_Making authorized? |
No — it is not inferred by elimination |
| Is task-stratified GIW validation authorized? | No |
First-party PerForm Lab / RIT structural evidence¶
A public Rochester Institute of Technology repository, PerForm-Lab-RIT/Pupil-Labs-Core-RITnet-Plugins, contains ritnet/Ellseg_v2/dataset_generation/ExtractedGIW.py. The script's in-file metadata identifies @author: rakshit and describes itself as extracting data from the Gaze-in-the-Wild project.
The reviewed source explicitly defines:
['Indoor_Walk', 'Ball_Catch', 'Visual_Search', 'Tea_Making']
and, for each task directory, addresses:
extracted_data/<task>/ProcessData_cleaned
It then loads each ProcessData, reads PrIdx and TrIdx, and uses those identities to address the raw recording tree at:
<path_data>/<PrIdx>/<TrIdx>/Gaze
This is first-party structural corroboration that the GIW processing code maintained task-separated extracted data while retaining participant/trial identities. It is materially stronger provenance than a secondary replication source.
The source is pinned to:
- repository:
PerForm-Lab-RIT/Pupil-Labs-Core-RITnet-Plugins - commit:
ebec5d1db118e39de60a14160f09ee33cd7f3b5d - tree:
67af6104f7d52c969be7737b04fa349be6a2ec4f - file blob:
ae51e50eb9f3d224417d1fd24253052196480870 - path:
ritnet/Ellseg_v2/dataset_generation/ExtractedGIW.py
Frozen GazeForge evidence:
validation/evidence/gaze-in-wild/gaze-in-wild-perform-lab-task-structure-corroboration-evidence-v1.json
Evidence fingerprint:
2352b5969285c4d3d182035570b3a3b76461fa683e50d69cf4ff0e2f6a827aec
Why the first-party evidence still does not close the numeric mapping gate¶
The reviewed script iterates the four named task directories and reads TrIdx from the ProcessData files inside each directory, but it does not itself enumerate a numeric lookup such as 1 → Indoor_Walk or 4 → Tea_Making. The code therefore verifies structure and identity linkage, not the complete distributed numeric mapping.
GazeForge deliberately does not turn list order, task-directory order, or the remaining unmatched publication task into a numeric mapping rule.
Secondary ACE-DNV numeric corroboration¶
The 2024 ACE-DNV study by Nejad et al. uses Gaze-in-the-Wild and explicitly describes the four activities as indoor walking, throwing/catching a ball, visual search, and making tea. Its public analysis repository contains participant/trial selection files that explicitly name three trial indices:
TrIdx |
ACE-DNV task label | Publication-normalized candidate | Observed participants in reviewed ACE-DNV files |
|---|---|---|---|
| 1 | Indoor_Walk |
indoor_navigation |
1, 2, 3, 6, 8, 10, 12, 17, 18, 22 |
| 2 | Ball_Catch |
ball_catching |
1, 2, 3, 6, 12, 16, 17 |
| 3 | Visual_Search |
visual_search |
8, 12, 19 |
| 4 | unresolved | unresolved | no explicit task-4 label in the reviewed ACE-DNV selection files |
The ACE-DNV evidence is pinned to repository commit 3142eb4457087743664d96994e952ed784741d1f and tree e52618478ba53f945b32f8056df297cf61c85f93. Exact reviewed file/blob identities are frozen in:
validation/evidence/gaze-in-wild/gaze-in-wild-ace-dnv-task-mapping-corroboration-evidence-v1.json
Evidence fingerprint:
e16aa3eea5c354ae0c6cb159cb3bdcaede19dfa535fdf23798f9fae3d35154b5
Paper DOI: 10.3758/s13428-024-02358-8.
Scientific boundary¶
The combined first-party structural evidence plus ACE-DNV numeric corroboration substantially narrows the open problem, but it does not authorize:
- a complete or authoritative
TrIdx→ publication-task mapping; TrIdx 4 → Tea_Makingby elimination;- task labels in the frozen exact participant-disjoint performance report;
- task-stratified Gaze-in-the-Wild validation;
- new empirical-performance claims;
- cross-dataset claims;
- native-60-Hz or Gazepoint GP3 validity;
- acquisition-hardware cadence claims;
- quarantine exit.
The next mapping gate is now narrower: recover a first-party distribution manifest, source file, task table, or equivalent authoritative record that explicitly binds the distributed numeric trial identities to publication tasks, including trial 4.