kfkas/ego-extra-imu-benchmark

Ego-EXTRA

Inference-friendly v2 manifest for kfkas/ego-extra-imu-benchmark.

This repository contains metadata and logical local asset references only. Full source MP4, VRS, CSV, NPZ, or ZIP files are not redistributed here.

Tables

table	rows	purpose
benchmark	14,644	one row per evaluation question
assets	149	one row per synchronized source asset
evidence	14,644	one row per question-to-asset time relation

The Viewer uses the benchmark split. The assets and evidence tables are separate Parquet/CSV files under data/v2/ because they intentionally have different columns. The common schema version is 2.0. sample_id joins benchmark to evidence, and asset_id joins evidence to assets.

Load for inference

from datasets import load_dataset

benchmark = load_dataset("kfkas/ego-extra-imu-benchmark", split="benchmark")
assets = load_dataset("kfkas/ego-extra-imu-benchmark", data_files={"train": "data/v2/assets.parquet"}, split="train")
evidence = load_dataset("kfkas/ego-extra-imu-benchmark", data_files={"train": "data/v2/evidence.parquet"}, split="train")

row = benchmark[0]
asset_rows = assets.filter(lambda item: item["asset_id"] in row["asset_ids"])
evidence_rows = evidence.filter(lambda item: item["sample_id"] == row["sample_id"])

video_relpath, imu_relpath, and frame_relpath are relative logical hints. Map them to the local sensor_vqa/raw/ root before opening media. query_time_raw and evidence_time_raw preserve benchmark-specific wall-clock or day-relative time.

CSV compatibility

CSV exports are available under data/v2/. List-valued fields are JSON strings in CSV and typed lists in Parquet.

Example

A short example clip is included for Viewer and visual sanity checks; it is not the complete source recording.

Sources

• https://fpv-iplab.github.io/Ego-EXTRA/

Follow the upstream license and access terms before downloading referenced media.

Five-minute development add-on (v1)

This is a derived development protocol, not the official Ego-EXTRA evaluation. Every logical input is one continuous 300-second interval from one source recording. The question is revealed only at clip-relative time 300 s; no video after that derived query endpoint is used. The clip can contain video after the original dialogue timestamp. It is therefore NOT an original-question causal-prefix score.

config	rows	status
dev5min_v1_candidates	570	question/dialogue interval inside clip; visual evidence not annotated
dev5min_v1_reviewed	4	coarse visual support checked by one assistant reviewer using sampled frames; smoke/development only

Candidate rows cover 91 source recordings and 536 logical clips. Exact duplicate QA at the same source time was removed locally. Dialogue-reference positions are spread across the input, but this is NOT a measured evidence-position distribution. Reviewed rows include an observed evidence_start_s/end_s and clip-relative equivalents. These are conservative support intervals, NOT frame-exact event/PNR ground truth. Unreviewed rows never receive invented evidence boundaries.

Loading and source QA join

Only newly derived scheduling metadata is added here: no QA text, choices, answers, transcripts, images, videos, sensor samples or native preparation code are copied in this update. Source QA remains in the unchanged historical default config and must be independently lawfully acquired. Its redistribution chain is not newly endorsed by this add-on. Join on question_id, preserving all five choices and the public answer without guessing a new answer.

from datasets import load_dataset

repo = "kfkas/ego-extra-imu-benchmark"
qa = load_dataset(repo, "default", split="benchmark")
schedule = load_dataset(repo, "dev5min_v1_reviewed", split="dev")
qa_by_id = {r["question_id"]: r for r in qa}

def add_qa(s):
    q = qa_by_id[s["question_id"]]
    result = {**s, **{k: q[k] for k in ("question", "choices", "answer_index", "answer_text")}}
    result["evaluation_choices"] = [q["choices"][i] for i in s["choice_order"]]
    result["evaluation_answer_index"] = s["choice_order"].index(q["answer_index"])
    return result

benchmark = schedule.map(add_qa)

Source QA was checked against revision ea6612149597d9c95932d96241e072816c485591 before this addition. Original answer-position skew is preserved in source fields; use the reproducible choice_order permutation and derived evaluation_choices/answer_index for evaluation instead. The add-on does not upload source answer labels. For temporally deictic questions ("now", "this part"), anchor the prompt to question_reference_start_s/end_s, or perform a separately reviewed retrospective rewrite; do not reinterpret the original question as about the final scene. An explicit timestamp anchor supplies a localization hint and must be reported when comparing retrieval systems. Ground-truth evidence fields are analysis-only and must not silently crop or bypass the full five-minute stream.

Local media and IMU

video_relpath and imu_relpath point to original locally acquired assets. Apply stream_start_s/end_s to both under the EXISTING relative-time join convention. The reviewed four local clips were also materialized without audio, with rebased IMU CSVs. They are NOT uploaded. IMU file presence and interval support were checked, but video-IMU clock offset is unmeasured, not asserted to be zero or calibrated. Thus this is a visual QA smoke set, not a sensor-timing qualification. At 1 FPS sample clip-relative ticks 0..299; at 0.5 FPS use 0,2,..,298. Always translate back to the source origin when reading original recordings.

Source terms and limitations

Official source: Ego-EXTRA. Release documentation states noncommercial restrictions and describes official 5/15/30-second clips before the question. General QA/media/sensor redistribution permission was not established. This add-on mirrors none of those payloads and grants no new rights to them. Obtain source assets through official routes and comply with their terms. No new upstream license is invented or assigned.

This is not a paper-ready held-out benchmark, a fine-event recall gold set, or an accuracy claim. Generic or expert-dependent questions can remain among candidates. Correlated questions/clips share a source_group_key; split by that key. The four reviewed examples are too small to support method-ranking claims. Citation: Ragusa et al., Ego-EXTRA: Video-Language Egocentric Dataset for EXpert-TRAinee assistance, WACV 2026, https://arxiv.org/abs/2512.13238.