Dataset Card for EgoHome

This is a FiftyOne dataset with 137 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("harpreetsahota/EgoHome")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

EgoHome is a benchmark of 137 human-annotated egocentric clips of household chores (about 2.0 hours at 1920×1080, 30 fps), cut from 46 longer videos. Each clip carries four kinds of annotation: hand boxes, subtasks, hand actions and hand captions, with segments as short as 0.10 s. This repository is the FiftyOne-format version of PerceptronAI/EgoHome: a video dataset in which the hand boxes are frame labels and the subtasks, hand actions and hand captions are temporal detections and temporal tags on the player timeline.

• Curated by: Perceptron AI
• Funded by: [More Information Needed]
• Shared by: harpreetsahota (FiftyOne-format conversion of the source dataset)
• Language(s): en
• License: cc-by-nc-4.0, with the source's additional permission: anyone, including commercial organizations, may use EgoHome to evaluate models and publish the results

Dataset Sources

• Repository: https://huggingface.co/datasets/PerceptronAI/EgoHome
• Paper: https://perceptron.inc/blog/egohome (blog post)
• Demo: [More Information Needed]

Uses

Direct Use

Evaluating models on dense, sub-second egocentric understanding of household chores: localizing the left and right hand, segmenting a clip into subtasks, and labelling what each hand does (an Ego4D verb class plus a free-text caption) over time. The source license explicitly permits using EgoHome to evaluate models and publish the results.

Out-of-Scope Use

Per the source license: selling the data, and training commercial models on it. The dataset has a single test split of 137 clips, so it is an evaluation benchmark and has no training or validation splits.

Dataset Structure

Topology

A FiftyOne video dataset (media_type: "video") with 137 samples, one per clip. There are no splits and no sample tags (the source has a single test split). Samples point at the original, unmodified source MP4s (H.264 High, 1920×1080, nominally 30 fps; 216,321 frames in total; 2.5 s to 62.0 s per clip; 6.33 GB). The 137 clips come from 46 source videos.

Fields

Sample fields:

Field	FiftyOne type	Description
id, tags, created_at, last_modified_at	standard	FiftyOne bookkeeping
filepath	StringField	Path to the clip MP4
metadata	VideoMetadata	Computed by FiftyOne from the video file: frame size, frame rate, frame count, duration, codec
clip_id	StringField	Clip identifier, e.g. 10c6450b22fd_004 (verbatim from source)
source_video_id	StringField	Id of the longer video the clip was cut from (verbatim from source)
clip_start_sec	FloatField	Offset of the clip inside its source video, seconds (verbatim from source)
subtasks	TemporalDetections	The clip's subtask timeline: 1,924 spans in total
left_hand_actions	TemporalDetections	Left-hand action spans, Ego4D verb classes: 2,374 in total
right_hand_actions	TemporalDetections	Right-hand action spans, Ego4D verb classes: 2,658 in total
left_hand_captions	TemporalDetections	Left-hand free-text caption spans: 2,520 in total
right_hand_captions	TemporalDetections	Right-hand free-text caption spans: 2,757 in total

Frame field:

Field	FiftyOne type	Description
frames.hand_boxes	Detections	One Detection per visible hand on each annotated frame: 417,542 in total (208,880 left, 208,662 right)

Attributes on each TemporalDetection (all four span fields and subtasks):

Attribute	Description
label	The span's text: a subtask, an Ego4D verb class, or a caption
support	[first, last] frame numbers (1-based, inclusive) of the frames inside the span
start_sec, end_sec	The span's exact start and end in seconds from clip start, verbatim from source
side	"left" or "right" for the per-hand fields; None on subtasks

Attributes on each hand Detection:

Attribute	Description
label	"left hand" or "right hand"
bounding_box	[x, y, w, h], relative to the frame
index	0 for the left hand, 1 for the right hand, so each hand is one track across frames
side	"left" or "right"
action	The hand's Ego4D verb class at that frame, or None if no action span covers it
caption	The hand's caption at that frame, or None if no caption span covers it
subtask	The clip's subtask at that frame, or None

Label types and why

• Detections per frame for hand boxes. The source stores one box per hand per annotated frame, so the boxes are frame labels, not sample labels. The hand's current action, caption and the clip's current subtask are copied onto each box as attributes, so boxes can be filtered by what the hand is doing.
• TemporalDetections for subtasks, actions and captions. The source stores each as {start, end, label} spans in seconds, which is what a temporal detection is. Left and right hand spans overlap in time, so each hand has its own action field and its own caption field instead of one shared list.
• Temporal tags as well. Every span is also stored as a temporal tag so it appears on the player timeline and can be scrubbed and searched: subtask: <text>, left hand: <verb> or right hand: <verb> (the head verb of the Ego4D class, e.g. take), and left caption: <text> or right caption: <text>. There are 12,233 tags, one per span (1,924 + 5,032 + 5,277).

dataset.info contents

dataset.info holds four notes: source (the source repository URL), license, frame_numbers (frame numbers are 1-based and a span covers the frames whose presentation time is in [start_sec, end_sec)), and boxes (how the source's normalized boxes were converted). dataset.classes lists the 47 distinct Ego4D verb classes that occur in this dataset for left_hand_actions and right_hand_actions.

Parsing decisions

• Frame times. 72 of the 137 clips have frame times that differ from an exact 30 fps grid by more than 1 ms, with gaps between consecutive frames up to 0.133 s. A span's support is therefore chosen by each frame's real presentation time, not by time × 30. A span shorter than one frame gap maps to the single frame at or after its start. start_sec and end_sec keep the exact source seconds.
• Box coordinates. The source stores boxes as [x_min, y_min, x_max, y_max] normalized to 0–1000. They are converted to FiftyOne's relative [x, y, w, h]. Every converted box lies inside the frame.
• Box timing. Each source box entry is matched to the video frame whose presentation time is within 1.1 ms of the entry's time. A hand that is null (off screen) gets no Detection on that frame. Frames with no box entry have no frame document.
• Dropped boxes. 209 box entries (370 hand boxes) are timed after the last video frame of their clip and are not included. The source card counts 417,912 boxes; this dataset holds 417,542. No subtask, action or caption span was dropped.
• Source duration. The source's duration_sec is approximate and is not stored: for example 10c6450b22fd_004 lists 60.0 s while its video runs 61.13 s. Use metadata.duration.
• Empty tracks. Three clips have no left-hand action spans.
• Source fields not stored as sample fields. fps, width, height and duration_sec are covered by metadata.

Dataset Creation

Curation Rationale

EgoHome was built to measure dense, sub-second egocentric understanding. The source card describes it as the first egocentric benchmark to feature four types of annotations per video and segments as short as 100 ms (773 of its segments are shorter than 0.5 s). Its metrics let predicted and ground-truth segments match many-to-many, because a chore can be divided into subtasks at several valid levels of detail.

Source Data

Data Collection and Processing

The 137 clips are cut from 46 longer household-chore videos, filmed with a head-mounted camera at 1920×1080 and 30 fps. Clip lengths range from 2.5 s to 62.0 s, and clip_start_sec gives each clip's offset in its source video. Further detail on how the source videos were selected and cut is [More Information Needed].

Who are the source data producers?

[More Information Needed]

Annotations

Annotation process

The source card describes the annotation as a synthetic seed that was human-corrected and then reviewed. It lists 417,912 hand boxes at 30 fps for the left and right hand, 1,924 subtask segments, 5,032 hand-action segments labelled with Ego4D verb classes, and 5,277 hand-caption segments. The source card gives a mean subtask length of 3.57 s and a mean segment length of 2.62 s across all segment types. The source card says its subtask metric was developed while finding a fair way to calibrate annotators; further detail on tools and guidelines is [More Information Needed].

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Citation

BibTeX:

@article{perceptron2026egohome,
  author = {PerceptronAI},
  title = {Perceptron EgoHome: A new benchmark for dense, sub-second egocentric understanding},
  journal = {PerceptronAI Research Blog},
  year = {2026},
  note = {https://perceptron.inc/blog/egohome},
}

APA:

PerceptronAI. (2026). Perceptron EgoHome: A new benchmark for dense, sub-second egocentric understanding. PerceptronAI Research Blog. https://perceptron.inc/blog/egohome

More Information

This dataset is a FiftyOne-format conversion of PerceptronAI/EgoHome, shared under the same CC BY-NC 4.0 license with attribution to Perceptron AI. Changes from the source: the annotations are converted to FiftyOne labels as described under Parsing decisions, 370 hand boxes timed after the end of their video are omitted, and the source's fps, width, height and duration_sec columns are replaced by metadata. The video files are the source MP4s, unmodified.

The evaluation metrics and the published results for the benchmark (centroid hit fraction for hand boxes, and permissible subsegment Jaccard for subtasks, hand actions and hand captions) are defined on the source dataset card.

Dataset Card Authors

harpreetsahota

Dataset Card Contact

[More Information Needed]