AEA-annotations / README.md
WitneyWW's picture
README: add pyramid2_v29_1108
b6c3cd9 verified
|
Raw History Blame Contribute Delete
8.08 kB
metadata
license: other
tags:
  - egocentric
  - aria
  - audio-visual
  - spatial-audio
pretty_name: Aria Everyday Activities annotations + SAVVY-Bench annotations

Aria Everyday Activities (AEA) annotations + SAVVY-Bench annotations

Four parts:

  • subset_solo20/ — 20 curated single-person examples with videos, raw VRS, 7-channel mic-array audio and full spatial annotations (see below).
  • aea_official/ — official per-recording annotations (no VRS) for 50 recordings.
  • savvy/ — SAVVY-Bench ground-truth annotations.
  • pyramid2_v29_1108/ — the last pyramid2 (v29) evaluation run on AEA: prompts, the evidence records the Examiner used (with per-record annotation source), run contracts and full run outputs, and AEA-specific code. Provisional, model-generated, not human-verified. See its README.

subset_solo20/

20 task clips (from 18 recordings) in which one person carries out a whole task alone, step by step, with complete spatial annotations and where spatial audio matters (a large share of the sound reaches the wearer from the side or from behind).

Selection

  1. Single person, whole task. Candidates are protocol-style clips cut from AEA single-person scripts (script 1/4/5). Gemini 3.1 Pro cut each recording into "one goal + ordered steps" segments; a second Gemini pass re-checked each clip alone. A clip is kept only if: no other person is present, little or no filler activity, steps happen in order, the end state is reached, protocol score ≥ 4, and at least max(4, #steps − 1) steps are visible. 84 → 44 clips.

  2. Complete spatial annotations. Every selected recording has the official raw VRS plus all official MPS outputs: closed/open-loop trajectory, online calibration, semi-dense points + observations, eye gaze, and speech.

  3. Spatial audio matters. For each clip we ran SRP-PHAT direction-of-arrival on the 7-mic Aria array (mic positions from the device calibration, directions in the CPF head frame). Checked against SAVVY-Bench partner-speech angles on 3 two-person recordings: median error 12°, 69% within 30°. Two measurements:

    • dense: fraction of loud 0.5 s windows whose sound arrives > 45° off the wearer's forward axis (frac_offaxis45) or > 90° (frac_behind90);
    • annotated: number of annotated non-speech sound events arriving > 45° off-axis.

    spatial_audio_score = frac_offaxis45 + 0.5·frac_behind90 + 0.05·n_annotated_offaxis. Top 20 kept, at most 2 clips per recording.

Note: in single-person recordings most sounds are made by the wearer's own hands, so they come from straight ahead. Roughly the top 8 examples have a strong spatial-audio signal (music, a vacuum or broom, doors or drawers to the side or behind). The rest only have 2–11% off-axis sound.

# example goal sec off-axis ≥45° behind ≥90° annotated off-axis events
1 loc3_script1_seq3_rec1__s0 Tidy the couch 40 0.40 0.30 5
2 loc2_script1_seq3_rec1__s1 sweep the living room floor 78 0.48 0.26 1
3 loc2_script5_seq1_rec1__s0 Bring luggage into the bedroom and open it 36 0.42 0.12 1
4 loc1_script1_seq3_rec1__s1 Wind the cord and put away the vacuum cleaner 41 0.20 0.05 2
5 loc4_script1_seq6_rec1__s1 Fetch drinks from the refrigerator 31 0.24 0.12 0
6 loc1_script5_seq3_rec1__s0 Unpack a suitcase 83 0.17 0.14 1
7 loc1_script5_seq3_rec1__s1 Put away a charger 14 0.18 0.09 1
8 loc1_script4_seq5_rec1__s0 serve food and sit at the table 23 0.18 0.06 1
9 loc4_script1_seq1_rec1__s0 do push-ups using handles 19 0.13 0.13 0
10 loc4_script5_seq7_rec1__s1 Update a grocery list 56 0.14 0.12 0
11 loc1_script1_seq6_rec1__s1 Serve a plate of pretzels 46 0.11 0.00 1
12 loc2_script5_seq6_rec1__s0 fold a pile of laundry 78 0.04 0.02 2
13 loc1_script4_seq4_rec1__s2 Season the eggs 44 0.07 0.03 1
14 loc1_script5_seq2_rec1__s1 prepare and start the washing machine 45 0.08 0.00 1
15 loc2_script1_seq6_rec1__s1 Fetch drinks from the fridge and place them on the table 30 0.10 0.05 0
16 loc2_script5_seq2_rec1__s1 Start a load of laundry 28 0.10 0.05 0
17 loc2_script1_seq6_rec1__s0 Portion pretzels into bowls 16 0.08 0.08 0
18 loc3_script5_seq1_rec1__s0 unpack items from a backpack 17 0.08 0.08 0
19 loc2_script5_seq3_rec1__s0 Move toiletries from the suitcase to the bathroom 12 0.11 0.00 0
20 loc2_script1_seq4_rec1__s0 inflate and tie a balloon 78 0.02 0.00 1

Layout

subset_solo20/
  index.json                         # all 20 examples: goal, steps, clip times (video sec + device sec), scores, per-event DOA
  examples/<seg_id>/
    clip.mp4                         # task clip, 456 px, 20 fps, stereo audio (cut from the recording video)
    mic_array_7ch_48k.wav            # same time span, 7-channel Aria mic array from the VRS (float32, peak-normalised)
    meta.json                        # same entry as in index.json
  recordings/<rec>/
    recording.vrs                    # official raw Aria recording (RGB, SLAM cams, eye cams, 7-ch audio, IMU ...)
    video_456px_20fps_stereo.mp4     # full undistorted recording video (from SAVVY-Bench release)
    official/                        # speech.csv, metadata.json, mps/slam/*, mps/eye_gaze/*  (official AEA)
    legacy/                          # narration.json, sound.json, traj_gaze_20fps.json  (SAVVY working files, NOT official)

Time alignment: clip_start_sec / clip_end_sec are seconds into video_456px_20fps_stereo.mp4. clip_*_device_sec are Aria device-time seconds, the same time base as the VRS, MPS and speech.csv (whose timestamps are in ns or µs). Legacy sound.json times are device seconds as well. Provenance: legacy/narration.json contains automatic per-second captions and legacy/sound.json contains sound-event labels; both come from the SAVVY project's working files. The goals and steps are Gemini drafts and have not been verified by a human.

aea_official/

Per-recording files from the official Aria Everyday Activities release, for 50 recordings (loc{L}_script{S}_seq{Q}_rec{R}):

  • speech.csv — speech-to-text words with startTime_ns, endTime_ns, written, confidence
  • metadata.json — recording metadata
  • mps/slam/ — MPS closed/open-loop trajectories, online calibration, semi-dense points, summary
  • mps/eye_gaze/ — MPS eye gaze

savvy/

Ground-truth annotation files from SAVVY / SAVVY-Bench (spatial audio-visual QA on AEA, University of Washington NeuroAI Lab: Mingfei Chen, Zijun Cui, Xiulong Liu et al.):

  • SAVVY-Bench/data/spatial_avqa/ — benchmark QA (savvy_bench.json, savvy_bench_file.json, gtavmap.json, human_eval_id.json)
  • SAVVY/data/test_json/ — test split json incl. GT AV map
  • SAVVY/data/vis_gt_keyframes_all/, SAVVY/data/speech_gt_keyframes_all/ — visual / speech GT keyframes per recording
  • SAVVY/data/dense_traj/, SAVVY/data/correct_traj/ — trajectories
  • LICENSE, README.md — original SAVVY repo license and readme

Audio tracks and model prediction outputs from the SAVVY repo are not included.

License

  • aea_official/, subset_solo20/ (VRS, official, video): subject to the Project Aria / Aria Everyday Activities dataset license terms.
  • savvy/: BSD Modified License, © University of Washington (see savvy/LICENSE).

Please cite the original works:

@article{lv2024aria,
  title   = {Aria Everyday Activities Dataset},
  author  = {Lv, Zhaoyang and Charron, Nicholas and Moulon, Pierre and others},
  journal = {arXiv preprint arXiv:2402.13349},
  year    = {2024}
}

SAVVY: Chen et al., "SAVVY: Spatial Awareness via Audio-Visual LLMs through Seeing and Hearing", NeurIPS 2025.