What Is HD-EPIC?

HD-EPIC is introduced in HD-EPIC: A Highly-Detailed Egocentric Video Dataset (arXiv 2502.04144), with an official project page and an official GitHub annotations repository. The abstract calls it a validation dataset. Read: the abstract, the project page and the repository README only.

HD-EPIC is third-party research. Firsthand is not affiliated with its authors and claims no use, training or compatibility with it or any named model. Firsthand captures first-person human video, fast, on demand and to a written spec: 21 keypoints per hand with a visibility flag, wrist trajectories and action segments, plus one task instruction per episode.

The abstract describes newly collected kitchen egocentric video from unscripted recordings in diverse home environments, and the project page calls them unscripted multi-day recordings. It states 41 hours in 9 kitchens, with annotations grounded in 3D through digital twins of the scene. The README’s intermediate-data section refers to per-video Aria glasses device calibration, per-frame camera pose and gaze. Camera specs, resolution, participant count and consent process are Not stated in the sources read.

The abstract lists recipe steps, fine-grained actions, ingredients with nutritional values, moving objects and audio, primed with gaze. The abstract counts 59K fine-grained actions, 51K audio events and 263 annotations per minute. The repository releases narrations as a pickle, audio as CSV and the rest as JSON.

The benchmark is visual question answering: 26K questions in the abstract, 26.6K on the project page, across seven categories from recipes to gaze. The abstract reports Gemini Pro at 38.5% and calls the benchmark difficult. Firsthand draws no comparison from that score. Scores for the action, sound and segmentation assessments are Not stated in the sources read.

HD-EPIC is a separate dataset from EPIC-KITCHENS, and the table keeps each source’s facts apart. The sizes differ: 41 hours here, 100 hours in EPIC-KITCHENS-100 and 3,670 hours in Ego4D, per their entries. This page ranks none of them. HD-EPIC licence terms are Not stated in the sources read.

The sources read describe HD-EPIC as a validation dataset and a question-answering benchmark, alongside recognition and segmentation. They do not describe robot actions or 3D hand pose keypoints; hand masks are pixel regions, not keypoints. Suitability for training a manipulation policy is Not stated in the sources read, and this page draws no conclusion. The scope is kitchens only.

Write tasks, environments and signals into a spec and confirm them on a Spec Pilot. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Recipe, nutrition, audio-event and gaze labels like HD-EPIC’s are not fields this page states for Firsthand capture, so plan them separately.

What is HD-EPIC?

A published dataset of unscripted kitchen egocentric video with interconnected labels grounded in 3D, from a CVPR 2025 paper.

How large is HD-EPIC?

The abstract states 41 hours of video in 9 kitchens, with 69 recipes and 59K fine-grained actions. The project page adds 4.4M frames.

What is the HD-EPIC benchmark?

A question-answering benchmark of 26K questions across seven categories. The abstract reports Gemini Pro at 38.5%.

Is HD-EPIC the same as EPIC-KITCHENS?

No. It is a separate dataset, and its project page lists EPIC-KITCHENS-100 among connected datasets.

Can HD-EPIC be used commercially?

Licence terms are Not stated in the sources read. Check the download terms on the project page first.

Does Firsthand use HD-EPIC, and can I request a custom dataset?

Firsthand is not affiliated with the HD-EPIC authors and claims no use, training or compatibility. You can request a custom dataset, confirmed first on a Spec Pilot.