EgoShift / README.md
FShim008's picture
EgoShift v1.0 - WACV 2027 public release
6fa68f9 verified
|
Raw History Blame Contribute Delete
5.68 kB
metadata
license: other
license_name: cc-by-4.0-annotations-apache-2.0-frames
license_link: LICENSE
pretty_name: EgoShift
tags:
  - vision-language
  - spatial-reasoning
  - benchmark
  - evaluation
  - egocentric
  - wacv2027
task_categories:
  - visual-question-answering
size_categories:
  - 10K<n<100K

EgoShift

Bounding the Action-Only Shortcut in Egocentric Spatial Updating for Vision–Language Models · WACV 2027, Evaluation & Datasets track

Faisal Ahmed, John Quarles, Kevin Desai — Department of Computer Science, University of Texas at San Antonio

Paper · Code · Project page

Three egocentric spatial-reasoning tasks over simulated indoor scenes, with the evidence behind every number in the paper.

task question held-out items
A Same object seen twice, or two different objects? 5,838
B After a stated in-place turn, where is the object? 1,098
C In the second view: visible, out of view, or occluded? 5,850

When a question states a camera turn, that turn alone rules out some answers, so a predictor given only the stated action already beats chance. EgoShift computes the highest score that shortcut can reach before any model is run, from the camera geometry, and builds every turn to sit on that ceiling. For Task B's held-out set the ceiling is 504/1,098 = 45.90%, and the build matches it in every one of the ten action cells.

Start here

egoshift_reviewer_bundle.tar.gz     169 MB, everything below
croissant.json                      Croissant + Responsible AI metadata
CHECKSUMS.json                      SHA-256 for every file here

tar xzf egoshift_reviewer_bundle.tar.gz && cd egoshift_reviewer_bundle

Layout

code/          construction, scoring and audit tools, plus tests
task_{a,b,c}/  questions for all three splits, dev-split answers,
               manifests, and a seeded 120-item frame sample
model_panel/   276,403 raw model completions across the 36 reported models
excluded_runs/ runs that are not part of the reported panel, with reasons
audits/        124 build and audit reports
structure/     held-out cell structure with all bearings removed
human/         task-level human aggregates only
licenses/      LICENSE and NOTICE for the shipped bytes

code/ inside the bundle is a frozen snapshot taken at submission. For a maintained, tested copy — the one CI runs against — use github.com/UTSA-VIRLab/EgoShift.

What you can check without the answer key

The structural target. structure/task_b_action_cell_counts.json holds the per-cell counts and the derived ceiling for all 1,098 Task-B items.

Answer distributions and the failure mode. The raw completions contain each model's parsed answer per item. structure/task_b_holdout_structure.jsonl gives the rotation and turn for each item, so the per-cell answer behaviour reported in the paper can be recomputed directly.

Scene disjointness. The same structure files carry scene_id and physical_scene_component_id for held-out items, so the claim that no scene component is shared across splits is checkable.

The full pipeline on dev. Development-split answers ship, so the scorer and audits can be run end to end and compared against the shipped audit reports. Against the maintained repository:

git clone https://github.com/UTSA-VIRLab/EgoShift.git
cd EgoShift && pip install -e ".[test,full]" && make check

Runs that are not in the panel

excluded_runs/ holds attempts that the paper does not count, each with a stated reason in EXCLUSIONS.json. All three are documented access or runtime failures whose files contain probe and error stubs rather than answers. They ship so that an unavailable model is visible as a recorded failure rather than a silent gap in the panel.

Licensing

Annotations, questions, metadata and audit outputs are CC-BY-4.0. Rendered frames derive from AI2-THOR and ProcTHOR and are redistributed under the upstream Apache-2.0 terms with attribution and the NOTICE file preserved. licenses/ in the reviewer bundle carries both. Held-out answers are withheld entirely.

What is deliberately withheld

Held-out answers for all three tasks. All bearing geometry for held-out items: the post-turn bearing determines the answer outright, and the source bearing plus the stated rotation recovers it, so neither travels. Human per-rater and per-item responses, the intake files, and the pseudonym map.

Large archives

The full rendered frame sets are separate and are not needed to reproduce the paper's tables: egoshift_task_a.tar.gz (5.9 GB), egoshift_task_b.tar.gz (573 MB), egoshift_task_c.tar.gz (3.9 GB), egoshift_surveys.tar.gz (299 MB, the blank annotator survey instruments).

Reporting results

Report accuracy per stated turn, not pooled. That is the failure this benchmark documents: pooled Task-B accuracy puts the best configuration at 83.6% against a trained-human 84.6%, which reads as nearly solved, while the same system trails the panel badly at 90° and 180°.

Citation

@inproceedings{ahmed2027egoshift,
  title     = {EgoShift: Bounding the Action-Only Shortcut in Egocentric
               Spatial Updating for Vision--Language Models},
  author    = {Ahmed, Faisal and Quarles, John and Desai, Kevin},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}