kfkas/ego3-qa-imu-links

EGO3 streaming QA — connection information only

This is a locator and protocol publication, not a bundled video/QA/IMU dataset. It contains no question text, no choices, no answers, no evidence windows, no transcripts, no video, no IMU samples and no features. It publishes the identifiers, source revisions, relative asset locators, streaming input intervals and verified IMU join facts needed to connect an official local copy of three egocentric benchmarks to a causal streaming evaluation.

You cannot score a model from this repository alone. Obtain the question and answer payloads and the media from the official sources below, under their own agreements, then join on question_id and source_id.

Contents

config	rows	sources	what it locates
egoschema_public500	500	500	public-answer EgoSchema subset, one 3-minute clip per question
qaego4d_mc_test	500	148	ReKV/GroundVQA QaEgo4D MC test clips
rvs_ego	1465	10	RVS egocentric long streams with official per-question query times
qaego4d_mc_test_imu_joins	44	44	verified Ego4D IMU alignment and coverage for QaEgo4D clips

Total located questions: 2465. No row was deleted for lacking IMU.

Published field whitelist

Question rows carry only: dataset_id, config, split, question_id, source_id, group_id, stream_order, source_row_index, stream_start_s, stream_end_s, query_time_s, raw_query_time_s, query_time_clamped, evaluation_context, time_basis, video_relpath, video_duration_s, video_fps, annotated_duration_s, media_present, modalities, source_repo, source_revision, source_url.

The IMU join table carries only identifiers, the clip interval inside its Ego4D parent, the relative sensor locator, clock/unit/axis facts and measured coverage and gap statistics.

A per-config whitelist and a forbidden-name check are enforced at build time, so an upstream column cannot silently widen this release.

video_relpath and imu_relpath are relative logical paths that resolve against a local preparation root. They are not host paths and not download URLs. media_present describes the preparing host's local copy, not this repo.

modalities lists only modalities that genuinely exist for a row. Video-only rows carry no IMU fields at all; nothing is padded with nulls or placeholders.

Streaming input protocol

config	admissible input	question time
egoschema_public500	full 3-minute clip from 0 s	after clip end (time_basis = derived from the official protocol)
qaego4d_mc_test	full designated QA clip from 0 s	at clip end (time_basis = derived from the official protocol)
rvs_ego	accumulated stream from 0 s, memory kept across questions in stream_order	official per-question query_time_s

stream_start_s, stream_end_s and query_time_s define the admissible input interval. No future frames. Annotated evidence windows are deliberately not published here: they are answer-adjacent telemetry and they must never be used to crop the input.

For rvs_ego, raw_query_time_s preserves the official value and query_time_clamped marks the rows where it was clamped to the probed media duration.

IMU coverage

IMU exists for qaego4d_mc_test only, and only partially. The 148 clips were resolved to 107 distinct Ego4D parent videos through the official ego4d.json; only 28 parents are flagged has_imu. Alignment was measured by one-second occupancy across each clip's real interval inside its parent, using the official canonical timestamps.

quantity	value
clips with an accepted IMU join	44
clips at >= 99% interval coverage	30
distinct parents behind those joins	28

Many official IMU rows carry empty or implausible canonical_timestamp_ms values; those are excluded rather than trusted, and parent_canonical_fraction records how much of each parent file was usable. Alignment was never certified from matching filenames, similar motion, the first sensor timestamp or duration alone. Two accepted joins stay honestly low (0.129 and 0.152 coverage, interior gaps near 410 s) and are reported rather than dropped.

egoschema_public500 and rvs_ego have no time-aligned IMU and are video-only.

Sources, revisions and acquisition

set	official source	revision used	terms as declared upstream
EgoSchema	https://egoschema.github.io/ via lmms-lab/egoschema	58350524ea7eb29c47000121f4f4b65eb6b4acb9	mirror card says MIT; the authors place the data under Ego4D terms
QaEgo4D MC test	https://github.com/Becomebright/ReKV, Becomebright/QAEgo4D-MC-test	not pinned locally	adaptation card says CC-BY-4.0; original QAEgo4D needs its own agreement in addition to Ego4D
RVS	https://huggingface.co/datasets/Becomebright/RVS	03232e8520da83a127686d5ebc04464197d57c41	CC-BY-NC-4.0: attribution, non-commercial
Ego4D recordings and IMU	https://ego4d-data.org/	n/a	signed Ego4D agreement required; media and sensors are not mirrored

See sources.json for the same provenance in machine-readable form. Cite Ego4D in addition to each benchmark. Publishing locators here grants no rights in the upstream payloads, and changing the serialization invents none.

Limitations

Local preparation checked media presence, readability, duration and FPS consistency and temporal admissibility. It says nothing about QA accuracy; no model, feature extraction or evaluation was run. Counts describe one prepared snapshot, not an official leaderboard split. The QaEgo4D local copy's exact upstream revision was not pinned at acquisition time and is reported as unknown rather than guessed.