CoolFace
Datasetpublic

paper-review-only/S-EMBER

S-EMBER S-EMBER is a benchmark for streaming episodic memory over long egocentric (wearable-camera) video. Given a video and a natural-language question about something that happened earlier in the recording, a model must recall the relevant moment and answer. This repository is an anonymized mirror provided for peer review. It contains the complete benchmark. Author, institution, and provenance information has been intentionally omitted for double-blind review.… See the full description on the dataset page: https://huggingface.co/datasets/paper-review-only/S-EMBER.

sourceHugging Facecc-by-nc-4.0updated 2mo agoView on Hugging Face
0likes464downloads
Dataset Card

S-EMBER

S-EMBER is a benchmark for streaming episodic memory over long egocentric (wearable-camera) video. Given a video and a natural-language question about something that happened earlier in the recording, a model must recall the relevant moment and answer.

This repository is an anonymized mirror provided for peer review. It contains the complete benchmark. Author, institution, and provenance information has been intentionally omitted for double-blind review.

Contents

PathDescription
sember_grounding.jsonlOpen-ended grounding QA (one line per question).
sember_mcq.jsonlThe same questions as 5-way multiple choice (one line per question).
videos/The source egocentric videos (.mp4), one per video_id.

Statistics

  • —9,429 questions over 3,134 videos.
  • —Each question is provided in both open-ended grounding and 5-way MCQ form.
  • —8 reasoning categories: time/duration, location trace, visual-detail recall, temporal ordering, counting objects/events, spatial-aware reasoning, object comparison, sequential action.

Question schema (MCQ)

Each line of sember_mcq.jsonl has:

question_id, video_id, video_category_broad, video_category, question, question_time, question_category, duration, options, correct_index, correct_letter, correct_option_source, ground_truths, answer_start_time, answer_end_time.

Privacy

The released videos are processed for privacy: faces are blurred and audio is removed prior to release.