EgoTextVQA-Indoor

Indoor split of EgoTextVQA, an egocentric scene-text VQA benchmark (Zhang et al., 2025). 2,216 open-ended questions over first-person indoor video clips that require reading text in the scene to answer.

Format

test.json is a list of records:

{
  "video_id": "0479bea8-d221-4c6a-8c91-60108e43fe56",
  "question_type": "book-related",
  "correct_answer": "the book is being placed on the shelf",
  "question": "where is the book titled fables?"
}

Videos (not included)

This repo hosts annotations only. Obtain the video clips (<video_id>.mp4, fps=6) from the original EgoTextVQA distribution, then either:

• place them under ~/.cache/lmms_eval/egotextvqa/videos/ (default), or
• set EGOTEXTVQA_VIDEO_DIR to point at the directory.

Filenames already match the video_id field (with .mp4 suffix); no rename needed.

Evaluation

EgoTextVQA-Indoor uses a GPT-based judge for semantic answer matching. Configure with JUDGE_OPENAI_API_KEY (and optionally JUDGE_OPENAI_API_BASE) to run the official metric. The model defaults to gpt-4o-mini (override with EGOTEXTVQA_GPT_MODEL).

Citation

@article{zhang2025egotextvqa,
  title={EgoTextVQA: Towards Egocentric Scene-Text Aware Video Question Answering},
  author={Zhang, Yuanhan and ...},
  year={2025}
}