kfkas/supermemory-vqa-clip-dev500-imu-benchmark

SuperMemory-VQA clip dev500 (RGB + IMU)

A development/test set for iterating on streaming-VLM methods. It is not a final benchmark and not an official SuperMemory-VQA split. It holds 500 four-choice questions with public answers, drawn from the SuperMemory-VQA Aria recordings (Persons 1-10), packaged as 165 logical clips of 60 recording sessions. Each clip has synchronized RGB video and IMU, and every question is asked at the end of its clip.

Example

One real row: question 1013 (Person_1, tier 5-10min, category visual_recall).

If the player does not load, open example_clip.mp4 directly.

Question: What brand was the dash cam I put in the closet earlier?

• A. The brand of the dash cam you put in the closet was Nextbase.
• B. The brand of the dash cam you put in the closet was AZDOME. (answer)
• C. The brand of the dash cam you put in the closet was a known brand.
• D. This question can not be answered.

field	value
evaluation clip	5-10min__Person_1_session_1_01312026_glasses_1266__00780, session seconds 780-1230, question asked at 1230 s
evidence	1056-1066 s, 164 s before the question (evidence_bin 4)
excerpt shown above	1052-1070 s of the same recording

The excerpt only illustrates the evidence; the benchmark input for this row is the whole 450 s clip, streamed together with its IMU. The excerpt is cut from the upstream recording, downscaled to 640 px, re-encoded at 15 FPS and stripped of audio. Faces and license plates are already blurred in the upstream release. It carries no IMU data.

Protocol

A clip is a logical interval [stream_start_s, stream_end_s] of one recording session. Stream it from stream_start_s at your chosen FPS together with the IMU of the same interval, then ask every question of the clip at stream_end_s. Rows that share source_group_key (the stream_id) share one stream, so compute scales with clip count, not question count. Never read frames after causal_input_end_s; future_frame_allowed is always false.

tier	questions	clips	clip length	stream hours at 1 FPS	evidence-to-question distance (median / p90)
5-10min	250	109	7.5 min	13.6	2.92 / 6.4 min
10-20min	250	56	15 min	14.0	7.0 / 13.0 min

How it was selected, in order:

1. Pool: answerable questions of Persons 1-10 whose evidence lies in a single session and whose evidence interval is usable. Unanswerable and multi-session questions are excluded.
2. Each question is placed in a 7.5-minute or 15-minute clip of its own session that contains all of its evidence; the clip end becomes the question time. 250 questions per tier.
3. The evidence start is spread evenly over the clip's one-minute positions (evidence_bin is the minute index of the evidence start inside the clip), so evidence is not always near the question.
4. Evidence reuse is minimized. SuperMemory-VQA writes several questions about one scene, so selection first uses only questions whose evidence interval does not intersect any selected question of the same session, and allows one more overlap at a time only when a bin cannot be filled. evidence_overlap_count is the number of other rows of the same session whose evidence intersects this one: 398 rows have 0, and the other 102 rows form 51 pairs (no row has more than 1 partner).
5. Persons are balanced as far as non-overlapping evidence allows: P1: 77, P10: 28, P2: 77, P3: 40, P4: 57, P5: 55, P6: 63, P7: 44, P8: 28, P9: 31.

Loading

``python from datasets import load_dataset

repo = "kfkas/supermemory-vqa-clip-dev500-imu-benchmark" bench = load_dataset(repo, "benchmark", split="dev") streams = load_dataset(repo, "streams", split="dev")

row = bench[0]

frames and IMU: row["video_relpath"] / row["imu_relpath"], from row["causal_input_start_s"]

to row["causal_input_end_s"] only; ask at row["effective_query_time_s"].

``

benchmark.jsonl and streams.jsonl are the line-oriented exports; the CSV files are compatibility exports that serialize list fields as JSON strings. benchmark rows join to streams on stream_id; source_ids[0] is the recording session.

Columns

benchmark uses the project's common flattened columns (dataset_id, question_id, source_group_key, choices, answer_index, answer_text, answer_key_status, video_relpath, video_start_s, video_end_s, query_time_s, raw_query_time_s, effective_query_time_s, causal_input_start_s, causal_input_end_s, evaluation_context, evidence_intervals, evidence_intervals_causal, required_modalities, modalities, native_fps, source_license, source_url, ...) so it loads like the other published benchmarks. This repository has real IMU, so it also carries imu_relpath, imu_relpaths, imu_format, imu_start_s, imu_end_s and source_time_offset_s. Development-specific columns: tier, person, original_split, stream_id, clip_length_s, evidence_start_s, evidence_end_s, evidence_offset_s, evidence_bin, evidence_distance_s, evidence_overlap_count.

• All times are seconds from the start of the session video, not from the clip start (except evidence_offset_s, which is clip-relative).
• evaluation_context is logical_clip_causal_at_clip_end: the input is the clip, the query is at its end.
• raw_query_time_s is null: SuperMemory-VQA question times belong to other recordings, so the query time here is derived (query_time_source = derived_clip_end_no_original_query).
• evidence_intervals is telemetry for analysis. It is not an input window and never permits reading outside the clip.
• Choices include the dataset's calibrated "This question can not be answered." option; only answerable questions are used, and answer_index is the public answer.

Media, IMU and local files

No full recording, VRS, IMU sample, gaze or audio is included; the only media is the short example excerpt shown above. video_relpath and imu_relpath are logical locators under your local sensor_vqa/raw/ tree; the part after sensor_vqa/raw/supermemory_vqa/ is the file path inside the upstream dataset (data/video/<Person>/<session>.mp4 and data/raw/<person>/<session>.vrs). Download only the 60 sessions listed in streams from the upstream repository under its terms.

• Video: 165 session videos were probed with ffprobe; 0 missing or corrupt; native FPS {'30.00': 165}.
• IMU: stored in the Aria .vrs files (streams imu-right, imu-left) and must be decoded locally. Video and VRS durations agree within about 0.15 s, but the IMU-to-video time origin has not been measured (imu_video_sync_status = duration_matched_origin_unmeasured; source_time_offset_s is the identity 0.0).

Limitations

• A derived within-session recall protocol with derived question times: scores are not comparable with published SuperMemory-VQA numbers.
• The questions cluster: 102 rows still share evidence with another row, and rows in one clip share the same stream, so the effective number of independent samples is below 500. As a rough rule of thumb (not a computed confidence interval), treat accuracy gaps below about 5 points as noise and confirm final claims on a full benchmark.
• Person balance is limited by where non-overlapping evidence exists (Persons 1 and 2 are larger).
• Public answer keys only; no private labels are included.

Provenance, license and redistribution

Source: SuperMemory-VQA, https://huggingface.co/datasets/OSU-AIoT-MLSys-Lab/SuperMemory-VQA (https://arxiv.org/abs/2606.00825); upstream card revision observed at publication: 1d228e0f10049a8a84c458dded2aa25b1e21ce8f. The upstream card declares CC BY-NC-SA 4.0. This repository redistributes annotation-derived rows (questions, choices, public answers, evidence times) under the same license: attribution required, non-commercial use only, share-alike. Changes made: subset selection, clip and query-time derivation, added columns. The example excerpt is an adaptation of upstream video (cut, downscaled, re-encoded, audio removed) shared under the same license. The recordings come from IRB-approved participant data; follow the upstream access terms and privacy rules before downloading media.

Generated 2026-10-09T16:29:05Z; common schema 3.0-imu; development set v3.