First-person video of one person's everyday life, recorded with Project Aria glasses from 2026-04-15 to 2026-08-18: 795.0 hours, 527 recordings, 114 recording days. Each recording has ~1 fps RGB frames (2880×2880), eye gaze, audio with transcript, dense captions, and a five-level activity hierarchy (L1 single actions → L5 major activities). NextAct is a 1,500-point benchmark for predicting what the person does next.

Project page: https://kkkkawayi.github.io/nextme-800/

raw/recordings.csv                         one row per recording: time, duration, frame counts, modalities, tar path
raw/YYYY-MM-DD/YYYY-MM-DD_HHMM-HHMM.tar    one recording (8.81 TB in total)
captions/1_nextme/<recording_id>/
    raw_caption.jsonl                      dense captions (530,839 segments)
    L1.jsonl … L5.jsonl                    activity hierarchy
captions/2_egolife/<participant>/L1…L5     the same hierarchy for the 6 EgoLife participants
nextact/                                   benchmark points and scripts
prompts/                                   caption and hierarchy prompts

recording_id = capture date and start–end time in Hong Kong time, e.g. 2026-04-15_1323-1709.

Raw recordings (`raw/`)

<tar_root_dir>/                                   (column tar_root_dir in recordings.csv)
├── picture/masked/<date>_<HH-MM-SS>_<TZ>__frame_<NNNNN>.jpg   RGB frame; faces and private on-screen text masked
├── picture/ocr_text/<same name>.txt                          text found in the frame
├── eye_tracking/<YYYYMMDD-HHMMSS-mmm>.jpg                     eye-camera image
├── eye_tracking/gaze.csv                                     gaze point in frame pixels for each frame
├── audio/anonymized.wav                                      48 kHz mono, voice-transformed
└── audio/transcript.txt                                      "[YYYY-MM-DD HH:MM:SS HKT -> ...] text"

2,878,042 RGB frames and 2,878,864 eye images in total; 11 recordings have no gaze.csv. Frame names carry the capture time in Hong Kong time; sort frames by frame_NNNNN (tar order is not chronological).

import tarfile
with tarfile.open("2026-04-15_1323-1709.tar") as t:
    frames = sorted(n for n in t.getnames() if "/picture/masked/" in n)
    jpeg = t.extractfile(frames[0]).read()

gaze.csv columns: frame_file, gaze_x, gaze_y (pixels), radius_px, in_bounds, depth_m, depth_ok, time_delta_ms, yaw_uncertainty_deg, pitch_uncertainty_deg, followed by the original Project Aria MPS fields.

Captions (`captions/`)

Captions were written by Gemini 3.7 Flash from 30-second windows (frames with gaze, gaze crops, OCR, transcript), in English; speech stays in its original language. Private strings (names, contacts, accounts, URLs, IDs) are replaced by XXX.

raw_caption.jsonl: recording_id, seg_id, start, end (YYYY-MM-DD HH:MM:SS), action (detailed first-person description), action_brief (= L1 text), objects, environment (scene, screen content, visible text, media, sound), text_visible, speech, details, approx_time (true for 55 segments whose exact boundaries are unknown).

L1.jsonl … L5.jsonl: recording_id, id, start, end, text. Each level groups the level below by goal; a parent's time span covers its children. Rows are in time order. Speech and actions can overlap in L1, and a higher-level event can contain a short interruption.

level	events	typical duration
L1	528,434	~5 s
L2	47,903	~1 min
L3	6,366	~7 min
L4	1,988	~24 min
L5	1,235	~39 min

EgoLife files use the same fields with participant and day (810,120 events, from the public EgoLife captions).

from datasets import load_dataset
l3 = load_dataset("mmm8383/NextMe-800", "nextme_L3", split="train")

NextAct benchmark (`nextact/`)

1,500 points: 1,000 from NextMe-800 and 500 from EgoLife, 300 per level. Given the previous events at one level, predict the next K events (K = 1 and 10 in the paper). points.jsonl stores subject, level, cutoff (position of the first ground-truth event on that subject's concatenated timeline) and target_windows.

cd nextact
python build_context.py --k 10 --n-context 50                  # build contexts and ground truth
python predict.py --model gpt-4o --k 10 --out preds_k10.jsonl  # any OpenAI-compatible model; 3 candidates per point
python predict.py --model repeat-last --k 10                   # baseline: repeat the last K events
python score.py --metric embedding --preds preds_k10.jsonl     # also: --metric reranker | llm-judge

Score: soft edit distance between predicted and true sequences (substitution cost = 1 − similarity), S = 1 − SED / max(m, n), best of 3 candidates. The embedding metric (Qwen3-Embedding-8B) is normalized to random predictions: 0 = chance, 1 = perfect.

License and citation

CC BY-NC 4.0. captions/2_egolife/ is derived from EgoLife and follows its S-Lab License 1.0 (captions/2_egolife/LICENSE_EgoLife.txt).

@misc{nextme800,
  title  = {NextMe-800: Anticipating Personal Behavior from Months of Egocentric Video},
  author = {Zhaoxu Meng and Yiming Sun and Mingyuan Gao and Jiachang Zhang and Zhuhan Dai and Yipeng Du and Zheng Lian and Jian-Qiao Zhu},
  year   = {2026}
}

Downloads last month: 887

Total file size:

8.81 TB

You need to agree to share your contact information to access this dataset

NextMe-800

Contents

Raw recordings (raw/)

Captions (captions/)

NextAct benchmark (nextact/)

License and citation

Raw recordings (`raw/`)

Captions (`captions/`)

NextAct benchmark (`nextact/`)