wearable-ai

Wearable AI Dataset (ECCV 2026) Part of the Wearable AI Workshop at ECCV 2026. A benchmark of egocentric (first-person, head-mounted wearable camera) videos paired with three complementary video question-answering tasks for evaluating wearable-AI assistants on real-world everyday activity videos. ▶ Baseline code & evaluation scripts: see starter_kit/README.md. The starter kit ships inside this repo, so git clone gives you the code and the data together. Tasks EgoLongQA — Long-form video question answering with multiple-choice answers. Given a video and a question with four options (A/B/C/D), predict the correct answer. EgoConv — Conversational video question answering. Given a video and a multi-turn conversation, generate free-form answers for each turn. The model sees only the video up to the current turn (no future leaking) and uses its own previous answers as context. EgoProactive — Proactive assistant over streaming egocentric video. Given a video stream, the model must decide when to speak and what to say at each candidate moment, simulating a wearable AI that volunteers helpful information without being prompted. Configurations Config Split samples Description egoconv val 700 Multi-turn conversational QA grounded in egocentric video egolongqa val 700 Long-video MCQ over single, longer egocentric clips egoproactive val 700 Proactive-assistant moments over streaming egocentric video Repo Layout facebook/wearable-ai/ ├── README.md # this file ├── LICENSE # CC-BY-NC-4.0 (dataset) ├── egoconv/ │ ├── wearable_ai_2026_egoconv_val_700.jsonl │ └── val/ .mp4 # 700 videos, ~91 GB ├── egolongqa/ │ ├── wearable_ai_2026_egolongqa_val_700.jsonl │ └── val/ .mp4 # 700 videos, ~203 GB ├── egoproactive/ │ ├── wearable_ai_2026_egoproactive_val_700.jsonl │ └── val/ .mp4 # 700 videos, ~23 GB └── starter_kit/ # baseline code + evaluation scripts ├── README.md ├── LICENSE # CC-BY-NC-4.0 (code) ├── run_evaluation.py └── ... Cloning & Disk Usage The full dataset is ~317 GB across 2,100