Skip to content

Hollywood2EM annotation provenance

GazeForge now freezes a third Hollywood2 provenance layer: author-level licensing declaration evidence for the later Hollywood2EM annotation distribution, together with upstream participant-ID context from the original Hollywood-2 data source.

This layer is deliberately separate from both the original Mathe–Sminchisescu gaze-data licence and the exact GIN repository-byte audit.

Author-level open-source declaration

Ioannis Agtzidis's 2020 TUM dissertation, Towards a better understanding of eye movements in natural contexts, introduces Chapter 4, Hand-labeled data sets, with the statement:

“All the data presented in this chapter are made publicly available with an open-source license.”

Footnote 2 on that statement points directly to:

https://gin.g-node.org/ioannis.agtzidis/hollywood2_em

The dissertation was submitted to the Technical University of Munich on 15 June 2020 and accepted on 8 September 2020. This is materially stronger than inferring rights from the Hollywood2EM article licence or from the licence of the older underlying gaze distribution: it is a dataset-author declaration that explicitly binds the chapter's Hollywood2EM data to the GIN repository.

The frozen evidence record is:

validation/evidence/hollywood2/hollywood2-annotation-provenance-evidence-v1.json

with fingerprint:

a08510e43caca2a8e6d5c85e7b1ad41c9f312247cd9bd8367372f8ecad8aacab

What the declaration does not prove

The dissertation does not name an exact licence identifier in the declaration, reproduce the licence text, or provide a repository LICENSE/COPYING file. GazeForge therefore keeps:

  • analysis_use_terms_status=unresolved;
  • raw_data_redistribution_terms_status=unresolved;
  • dataset_specific_license_verified=false;
  • license_inference_permitted=false.

The phrase “open-source license” is evidence that an author intended the Hollywood2EM data to be openly licensed; it is not substituted for exact legal terms. The article's CC BY 4.0 licence is still not treated as a dataset licence, and the original Hollywood-2 academic-use licence is still not inherited by the later GIN annotation repository.

Upstream subject-ID context

Stefan Mathe's 2015 dissertation, Actions in the Eye, provides a separate participant provenance fact for the original public Hollywood-2 eye-movement data: it states that the public dataset lists unique subject identifiers within each task group. The same chapter reports 12 Hollywood-2 action-recognition participants and 4 free-viewing participants.

The audited GIN annotation tree exposes 16 recurring filename tokens:

001 002 003 004 005 006 008 010 011 012 013 014 015 017 018 019

The matching count and the upstream unique-ID statement strengthen the case that the tokens are participant-like identifiers. They still do not establish an authoritative linkage between each GIN prefix and the original subject IDs or task groups. GazeForge therefore continues to record:

  • participant_identity_mapping_verified=false;
  • participant_group_membership_by_gin_token_verified=false;
  • mapping_inference_permitted=false.

Participant-disjoint modelling remains blocked.

Scientific boundary

This tranche verifies two previously missing provenance statements without crossing their evidential limits:

  1. the Hollywood2EM dataset author explicitly described the chapter data, including Hollywood2EM via footnote 2, as publicly available under an open-source licence;
  2. the original Hollywood-2 public dataset used unique subject IDs within groups.

It does not create an exact annotation-repository licence, redistribution permission, GIN-token-to-participant mapping, independent human-human agreement, participant-held-out model result, Lund↔Hollywood2 result, or canonical Frozen Evidence performance claim.

Next required evidence

The remaining high-leverage Hollywood2 gates are now narrower:

  1. recover exact Hollywood2EM licence text or an author/institutional statement naming the licence and redistribution scope;
  2. recover an authoritative statement or archive structure that links the GIN filename prefixes to original subject IDs and task groups;
  3. only after item 2, execute participant-disjoint Hollywood2 model validation;
  4. only after the same provenance boundary is satisfied, execute Lund↔Hollywood2 cross-dataset validation.