EgoActionBench: Benchmarking Dense Action Captioning for Egocentric Human Videos

Paper · Project page · GitHub

EgoActionBench is a benchmark for dense action captioning in egocentric human videos. Given a complete video clip, a model predicts what the camera wearer does, when each action occurs, and whether it involves the left hand, right hand, or both hands.

The dataset contains 1,407 videos and 41,132 English action annotations. Each action has a hand label, a temporal interval, a caption, and structured references for its verb, objects, tools, and motion. The MP4 files occupy 23.69 GB.

Files

EgoActionBench/
├── README.md
├── annotations.jsonl
├── metadata.jsonl
└── videos/
    ├── EAB_000001.mp4
    ├── EAB_000002.mp4
    ├── ...
    └── EAB_001407.mp4

Both JSONL files contain one record per video. Join them using video_id. Video IDs are fixed and match the MP4 filenames.

Download and load

Access currently requires approval on this dataset page. After approval, sign in with the same Hugging Face account and download the dataset:

hf auth login
hf download play0718/EgoActionBench --repo-type dataset --local-dir EgoActionBench

To download only the annotation and metadata files:

hf download play0718/EgoActionBench annotations.jsonl metadata.jsonl \
  --repo-type dataset --local-dir EgoActionBench

Read the downloaded files with Python:

import json
from pathlib import Path

root = Path("EgoActionBench")
with (root / "annotations.jsonl").open(encoding="utf-8") as f:
    annotations = [json.loads(line) for line in f]
with (root / "metadata.jsonl").open(encoding="utf-8") as f:
    metadata = {row["video_id"]: row for row in map(json.loads, f)}

sample = annotations[0]
video_path = root / metadata[sample["video_id"]]["file_name"]
actions = sample["gt"]

The two tables are also exposed as separate Hugging Face Datasets configurations:

from datasets import load_dataset

annotations = load_dataset("play0718/EgoActionBench", "annotations", split="test", token=True)
metadata = load_dataset("play0718/EgoActionBench", "metadata", split="test", token=True)

The test split contains the complete benchmark in each configuration. MP4 files can be downloaded separately with the commands above.

Annotations

Each line in annotations.jsonl contains:

Field	Type	Description
video_id	string	Video identifier, such as EAB_000001.
duration	number	Video duration in seconds used to define valid annotation and prediction intervals.
gt	array	Reference action annotations for the video.

Each action in gt contains:

Field	Type	Description
video_id	string	Identifier of the containing video.
gt_id	string	Action identifier, unique within the video. The prefixes l_, r_, and bh_ denote left, right, and both hands.
hand	string	left, right, or both.
start	number	Action start time in seconds from the beginning of the clip.
end	number	Action end time in seconds from the beginning of the clip.
caption	string	English description of the action.
facet	object	Structured references for verb, objects, tools, and motion.

Temporal intervals are half-open: [start, end), with 0 <= start < end <= duration. The published times are in seconds. Annotations from different hand tracks may overlap; use each action's explicit interval and hand label.

Example from EAB_000001, with one action shown:

{
  "video_id": "EAB_000001",
  "duration": 20.2,
  "gt": [
    {
      "video_id": "EAB_000001",
      "gt_id": "bh_1",
      "hand": "both",
      "start": 2.6,
      "end": 2.833,
      "caption": "Pass the domino from one hand to the other.",
      "facet": {
        "verb": {"status": "required", "value": "pass"},
        "objects": {
          "status": "required",
          "items": [{"item_id": "o1", "role": "patient", "value": "domino"}]
        },
        "tools": {"status": "not_applicable", "items": []},
        "motion": {"status": "required", "value": "translate"}
      }
    }
  ]
}

Structured references

Each facet has a status:

• required: the field has an annotated reference to evaluate.
• not_applicable: the field does not apply to this action.
• skip: the field is excluded from semantic scoring for this action.

verb and motion use a value when one is provided. Motion values are reach, translate, lift, lower, rotate, press, oscillate, and static.

objects and tools use an items array. Each item has an item_id, a semantic role, and a reference value. Item IDs are local to the corresponding action and facet. Fields marked not_applicable or skip are excluded from semantic scoring; their status should be preserved when reading the annotations.

Video metadata

Each line in metadata.jsonl contains:

Field	Type	Description
video_id	string	Identifier shared with annotations.jsonl.
file_name	string	Video path relative to the repository root.
source_dataset	string	egoschema, egolife, or epic.
fps	number	Average frame rate reported by the video stream.
num_frames	integer	Number of frames in the video stream.
width	integer	Video frame width in pixels.
height	integer	Video frame height in pixels.

{"video_id":"EAB_000001","file_name":"videos/EAB_000001.mp4","source_dataset":"egoschema","fps":30.0,"num_frames":606,"width":960,"height":540}

Evaluation

EgoActionBench evaluates semantic fidelity, temporal localization, and hand-action association. Temporally eligible predictions are compared with each reference action, and a prediction is selected by content agreement. Semantic fidelity, temporal overlap, and action precision and recall contribute to EgoActionScore.

The public evaluation toolkit is being prepared in the official repository. See its evaluation overview for the scoring framework.

Data sources

The videos originate from Ego4D, through EgoSchema, EgoLife, and EPIC-KITCHENS-100. Source videos remain subject to their original licenses and terms of use.

Citation

@misc{shen2026egoactionbench,
  title         = {EgoActionBench: Benchmarking Dense Action Captioning for Egocentric Human Videos},
  author        = {Shen, Yujiao and Song, Jiahe and Li, Haitian and Yang, Jingkang and Liu, Ziwei},
  year          = {2026},
  eprint        = {2611.03646},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2611.03646}
}