Public dataset comparison
Best Public Egocentric Video Datasets
The best public egocentric video dataset depends on the task, modality, domain, documentation, and licensing need. Ego4D, Ego-Exo4D, and EPIC-KITCHENS are public benchmark references, not commercial supply; teams must verify rights, access terms, consent posture, and use constraints before commercial use.
Verdict by buyer scenario
How we selected and evaluated the options
How this shortlist is built. These are public or academically accessible egocentric corpora with a primary source, compared by fit — not scored into a ranking. Weights are deliberately not applied: the right dataset is a function of your task and your rights posture, not a leaderboard.
| Criterion | Why it matters | Evidence we require |
|---|---|---|
| Task fit | A corpus built for kitchen action recognition rarely transfers to warehouse manipulation. | Official paper/project description of the task and domain. |
| Modality & annotations | Gaze, IMU, depth, and 3D hand tracking are present in some corpora and absent in others. | Documented modality/annotation list on the project or paper. |
| Public availability | A portal or gated application existing says nothing about rights. | A working access path (download, portal, or signed agreement). |
| Repository / license signal | A code or card license does not license the captured media. | The exact license file or dataset-card license statement. |
| Consent / privacy signal | Egocentric footage records homes, faces, and bystanders. | Any stated participant consent / anonymization; otherwise marked 'not verified'. |
| Commercial-suitability posture | Only a risk label, never a legal clearance. | Derived from the license + consent evidence, with a checked date. |
No weights are applied — this is a shortlist by task and rights posture, not a scored ranking.
- Inclusion rules
- Included if the corpus is public or academically accessible AND has a primary source (paper, project page, or dataset card) we can cite and date.
- Exclusion rules
- Excluded: private-only corpora, vendor marketing pages with no dataset access, synthetic-only resources unless labeled, and anything with no primary source. We do not list a dataset just because it ranks for the keyword.
- Source basis
- Primary papers, official project pages, dataset cards, and license files — each with a checked date. Numbers are mirrored from the versioned egocentric dataset registry that also backs this page's downloadable CSV/JSON.
- Update cadence
- Dataset terms and cards change; this shortlist is re-checked quarterly, or sooner when a release changes its license or access path.
- Disclosure
- truelabel publishes this page and offers custom rights-cleared egocentric collection — a commercial interest you should read this knowing. Public datasets are assessed from official public documentation, not paid placement, and truelabel is listed only as a custom-collection alternative, never as a public dataset. This is not legal advice; commercial suitability requires review of the exact release, license, consent, and intended use.
- Scoring caveat
- This is a shortlist, not a ranking. Terms change — trust the checked dates, and re-open the official source before anything load-bearing.
Evidence matrix
| Option | Supported claim | Official source | Checked | Confidence | Limitation |
|---|---|---|---|---|---|
| Public egocentric datasets | |||||
| Ego4D | Broad first-person daily-activity benchmark corpus (~3,670 hours of video from 931 unique camera wearers across 74 locations in 9 countries, per the paper abstract; the official project page and the egocentric registry list 923 participants); access is governed by a signed dataset agreement, not open download. | Ego4D: Around the World in 3,000 Hours of Egocentric Video | 2026-07-21 | High (paper) | No robot action labels; commercial use requires legal review of the executed agreement and per-modality coverage. |
| Ego-Exo4D | Paired egocentric + exocentric skilled-activity dataset (~1,286 hours, 740 participants, 13 cities, 123 sites) with gaze, IMU, point clouds, and camera pose. | Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives | 2026-07-21 | High (paper) | Access requires accepting published terms; no commercial right is inferred from access to code, annotations, or data. |
| EPIC-KITCHENS-100 | Kitchen-only egocentric benchmark (100 hours, 45 kitchens, 20M frames); public annotation materials are CC BY-NC 4.0. | Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-KITCHENS-100 · EPIC-KITCHENS-100 annotations license | 2026-07-21 · 2026-07-21 | High (paper + license) | The public CC BY-NC license does not permit commercial use; a separate commercial rights path must be requested from the team. |
| Egocentric-10K | Factory-floor worker-worn video (~10,000 hours; the dataset card's directory structure spans factories 001–085); the dataset card states Apache-2.0 for the release. | Egocentric-10K · Egocentric-10K dataset card and license | 2026-07-21 · 2026-07-21 | Medium (dataset card) | A stated data license does not by itself establish contributor consent, privacy, publicity, or third-party-content rights; the release is raw video without robotics annotations. The card no longer states worker or site totals — only the factories 001–085 structure. |
| EgoDex | Dexterous tabletop manipulation from Apple Vision Pro (~829 hours, 338,000 episodes) with 3D hand/finger tracking. | EgoDex: code and dataset release · EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video | 2026-07-06 · 2026-07-06 | High (project + paper) | The registered release terms are noncommercial, no-derivatives — not commercial clearance; code availability is not a grant over the data or recorded people. |
| HD-EPIC | Fine-grained kitchen egocentric corpus (~41 hours, 9 kitchens) with 3D digital twins and action annotations; EPIC-family extension. | HD-EPIC: A Highly-Detailed Egocentric Video Dataset | 2026-07-21 | Medium (paper) | The registered primary source is the paper; public data access and license terms must be confirmed separately before use. |
| Custom rights-cleared collection (not a public dataset) | |||||
| truelabel custom egocentric | A physical AI data marketplace that sources rights-cleared, task-matched egocentric capture with contributor consent artifacts and per-trajectory provenance — not a public dataset and not a reseller of one. | truelabel egocentric data licensing hub | 2026-07-19 | first-party | We publish this page; ask for relevant sample evidence, consent artifacts, and capacity before scale. |
Every public row above is also a row in this page's downloadable dataset registry (CSV/JSON). Public availability, repository/license terms, contributor consent, and commercial suitability are four different questions — a permissive card license (Apache-2.0) still leaves consent and privacy open; a noncommercial license (CC BY-NC, CC BY-NC-ND) closes commercial use outright.
Buyer decision checklist
- Choose when
- Benchmarking, pretraining, or scoping a capture plan → a public corpus that matches your task is enough. · Broad first-person activity → Ego4D; paired ego/exo skill → Ego-Exo4D; kitchen HOI → EPIC-KITCHENS-100; permissive raw volume → Egocentric-10K. · You need rights-cleared, task-specific footage with consent artifacts and QA tied to a deployment → custom collection, not a public dataset.
- Avoid when
- You need commercial-use rights from a CC BY-NC / CC BY-NC-ND release (blocked without a separate license), or your exact objects/environment/embodiment are absent from every public corpus.
- Proof to request
- The exact license file, a participant-consent / anonymization statement, the access process, and a sample you can load — for any public corpus before training, and from any custom supplier before scale.
Limitations and caveats
Comparison
| Dataset | Best public reference for | Primary limitation to evaluate |
|---|---|---|
| Ego4D | Broad egocentric daily-life activity research | Whether the task, access terms, and annotations fit your use |
| Ego-Exo4D | Skilled activity with first- and third-person perspective | Whether ego-exo capture maps to your sensor plan |
| EPIC-KITCHENS | Kitchen activity recognition and verb-noun action structure | Non-commercial licensing caveats and domain specificity |
Public egocentric dataset evidence registry
Source-checked scale, modality, and rights-review fields. A public code or data license does not by itself establish consent or every right needed for commercial model training.
| Dataset | Scale | Viewpoint and device | Modalities and domain | License and commercial-use review | Source |
|---|---|---|---|---|---|
| Ego4D | 3,670 hours — 923 participants — 74 sites | Egocentric Multiple head-mounted camera rigs | RGB video, audio, IMU (subset), gaze (subset) Daily-life activity and egocentric perception benchmarks | License: needs-review Commercial use: needs-review The access agreement, dataset components, and intended use need legal review. | Primary source License or access terms |
| Ego-Exo4D | 1,286 hours — 740 participants — 123 sites | Paired egocentric and exocentric Head-mounted and synchronized external cameras | RGB video, audio, gaze, IMU, 3D point clouds, camera pose Skilled human activities across multiple domains | License: needs-review Commercial use: needs-review No commercial right is inferred from access to code, annotations, or data. | Primary source License or access terms |
| EPIC-KITCHENS-100 | 100 hours — — 45 kitchens — | Egocentric Head-mounted RGB cameras | RGB video, audio, action annotations Kitchen activity recognition and anticipation | License: public Commercial use: prohibited The cited public license does not permit commercial use; ask the project about a separate rights path. | Primary source License or access terms |
| Egocentric-10K | 10,000 hours — — — — | Egocentric Worker-worn cameras | RGB video Factory and manual-work video | License: public Commercial use: needs-review A stated data license alone does not establish consent, privacy, publicity, or third-party content rights. | Primary source License or access terms |
| Egocentric-100K | 100,405 hours 2,010,759 clips — — — | Egocentric Worker-worn cameras | RGB video Large-scale manual-work video | License: public Commercial use: needs-review A stated data license alone does not establish consent, privacy, publicity, or third-party content rights. | Primary source License or access terms |
| EgoDex | 829 hours — — — — | Egocentric Apple Vision Pro | RGB video, 3D hand and finger tracking Dexterous tabletop manipulation | License: not-public Commercial use: prohibited The registered release terms are non-commercial and no-derivatives. | Primary source License or access terms |
| HD-EPIC | 41 hours — — 9 kitchens — | Egocentric High-resolution head-mounted cameras | RGB video, audio, 3D digital twins, action annotations Fine-grained kitchen activities | License: needs-review Commercial use: needs-review No commercial-use conclusion is drawn from paper or code availability. | Primary source |
| EgoVid-5M | — 5,000,000 clips — — — | Egocentric Mixed source video | RGB video, kinematic action labels, text action labels Egocentric video generation | License: needs-review Commercial use: needs-review Source-video rights, dataset terms, and intended commercial use require review. | Primary source |
Selection criteria (best-for, not best-overall)
Best means best for a stated evaluation need, not universally superior. This selection compares Ego4D, Ego-Exo4D, and EPIC-KITCHENS against source-backed criteria: task coverage, modality, documentation, annotation structure, access terms, and known limitations [1] [2] [3]. Each dataset answers different research questions [4] [5] [6].
When public datasets are enough
Public datasets can be enough for terminology alignment, early experiments, benchmark comparisons, and scoping a capture plan. They are less likely to be enough when a buyer needs rights-cleared, task-specific, environment-specific, or current data with consent artifacts and delivery metadata.
When custom or commercial data collection is needed
Custom collection is usually needed when a team requires rights-cleared task-specific footage, current devices and workflows, consent artifacts, retention rules, delivery metadata, or QA tied to a deployment environment. Treat public datasets as benchmark and research references unless access, consent, license, and commercial-use rights are separately reviewed [1] [2] [7].
Licensing and availability caveats
Every public dataset has its own access and license posture. EPIC-KITCHENS-100 annotation materials visibly state non-commercial restrictions, so commercial teams should treat it as a benchmark reference unless a separate approved rights path exists [7].
Ranking methodology for public egocentric datasets
This page ranks public references by task fit, viewpoint, annotation depth, documentation quality, access path, license clarity, and robotics relevance. It does not rank them by commercial usability: every production use still needs official license, consent, data-use, and redistribution review.
Fit by task: public reference vs custom-data gap
Public datasets are most useful when they match the observable task. For production robotics, the gap is often not visual quality; it is missing commercial rights, target-domain coverage, action/state streams, or consent artifacts.
| Task | Public starting point | Custom-data gap to check |
|---|---|---|
| Daily-life activity | Ego4D | Deployment location, wearer population, and license/use terms |
| Paired ego-exo skilled activity | Ego-Exo4D | Camera rig, synchronization, and domain transfer |
| Kitchen actions | EPIC-KITCHENS | Non-commercial caveat and target appliance/object coverage |
| Hand-object interaction | Ego4D / EPIC / HOI references | Contact labels, object state, failures, and robot action/state |
| Commercial VLA data | Public datasets for schema only | Rights-cleared robot episodes with language/action alignment |
Dataset cards buyers should keep beside the shortlist
A good public-dataset shortlist is a card set, not a winner-take-all ranking. Record the official source, modalities, annotation depth, access process, license terms, known domain skew, and the exact production question the dataset answers. Then decide what must be replaced by custom sourcing before training or evaluation.
| Dataset | Best use | Caveat before production |
|---|---|---|
| Ego4D | broad daily-life first-person pretraining and task vocabulary | review access terms, consent posture, and target-domain mismatch |
| Ego-Exo4D | paired skilled-activity viewpoint design | sync richness may not match simple single-view needs |
| EPIC-KITCHENS | kitchen verb/noun/action labels | non-commercial and home-context caveats need official review |
| HOI-oriented sets | hand pose, contact, object-state schema ideas | usually need license and capture-setting review |
| Custom capture | rights-reviewed target environment and delivery format | requires a clear pilot acceptance rubric |
License and access caveat box
Separate public availability from commercial permission. Check dataset access, annotation license, contributor consent, redistribution limits, derivative/model-use terms, and citation requirements. If any field is unclear, treat the public dataset as a benchmark/reference, run the issue through egocentric data licensing, and request rights-reviewed custom capture for production work through the robotics data marketplace.
When custom capture is the honest answer
Custom capture is not automatically better; it is better when the public set fails a specific gate. Common gates are target environment, current hardware, worker or bystander artifacts, task taxonomy, object distribution, action/state streams, and delivery format. A buyer should write the failed gate beside each public dataset in the shortlist, then decide whether to commission a small pilot through sourcing intake or keep the public data only for benchmarking.
Related pages
Use these to move from category-level context into specific task, dataset, format, and comparison detail.
External references and source context
- Egocentric video remains useful but incomplete for robot data buyers
Ego4D is an official public reference for egocentric video dataset scope, access, and dataset documentation.
ego4d-data.org ↩ - Ego-Exo4D project site
Ego-Exo4D is the official project source for paired first-person and third-person skilled-activity capture.
ego-exo4d-data.org ↩ - EPIC-KITCHENS project site
EPIC-KITCHENS is an official project reference for egocentric kitchen-activity data.
epic-kitchens.github.io ↩ - Ego4D: Around the World in 3,000 Hours of Egocentric Video
The Ego4D paper is the source-backed reference for first-person daily-life activity video and benchmark design.
arXiv ↩ - Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives
The Ego-Exo4D paper describes skilled human activity from first- and third-person perspectives.
arXiv ↩ - Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-KITCHENS-100
The EPIC-KITCHENS-100 paper supports public kitchen-activity benchmark facts and caveats.
arXiv ↩ - EPIC-KITCHENS-100 annotations license
The EPIC-KITCHENS-100 annotation license is a visible source for non-commercial licensing caveats.
GitHub ↩ - Ego-Exo4D annotations documentation
Ego-Exo4D annotation documentation supports dataset-structure and skilled-activity-label discussion.
docs.ego-exo4d-data.org - truelabel egocentric data glossary
Internal contextual link to the egocentric data definition.
truelabel.ai - truelabel egocentric warehouse video sourcing spec
Internal contextual link to warehouse egocentric video sourcing.
truelabel.ai - truelabel egocentric kitchen video sourcing spec
Internal contextual link to kitchen egocentric video sourcing.
truelabel.ai - truelabel industrial egocentric video sourcing spec
Internal contextual link to industrial egocentric video sourcing.
truelabel.ai - truelabel VLA training data sourcing
Internal contextual link to VLA training data sourcing.
truelabel.ai - truelabel warehouse robotics data sourcing
Internal contextual link to warehouse robotics data sourcing.
truelabel.ai - truelabel kitchen manipulation data sourcing
Internal contextual link to kitchen manipulation data sourcing.
truelabel.ai - truelabel LeRobot format guide
Internal contextual link to the LeRobot format guide.
truelabel.ai - truelabel LeRobot dataset alternative comparison
Internal contextual link to the LeRobot dataset alternative comparison.
truelabel.ai - truelabel eval data for robotics hub
Internal contextual link to robotics eval data sourcing.
truelabel.ai - truelabel teleoperation training-data page
Internal contextual link to teleoperation training data sourcing.
truelabel.ai - truelabel robot demonstrations training-data page
Internal contextual link to robot demonstration training data sourcing.
truelabel.ai - truelabel hand-object interaction data page
Internal contextual link to hand-object interaction training data requirements.
truelabel.ai - truelabel egocentric video datasets hub
Internal contextual link to the egocentric video datasets hub.
truelabel.ai
FAQ
What are the best public egocentric video datasets?
Ego4D, Ego-Exo4D, and EPIC-KITCHENS are common public references, but best depends on task, modality, domain, documentation, and license needs.
What is Ego4D used for?
Ego4D is used as a broad research reference for first-person daily-life activities and egocentric-video benchmark design.
What is EPIC-KITCHENS used for?
EPIC-KITCHENS is used as a public egocentric kitchen-activity benchmark, especially for action-recognition framing.
What is Ego-Exo4D used for?
Ego-Exo4D is useful when teams need to reason about skilled activity from both first-person and third-person perspectives.
Are public egocentric datasets enough for robotics models?
They can help with benchmarks and scoping, but production robotics teams often need task-specific capture, rights review, consent artifacts, and QA tied to deployment needs.
Which public egocentric dataset is best for kitchen tasks?
EPIC-KITCHENS is the strongest named kitchen benchmark, but buyers must review official license/access terms and should source custom data when commercial rights or target appliances/layouts are required.
Which public egocentric dataset is best for paired first- and third-person activity?
Ego-Exo4D is the main public reference for paired egocentric and exocentric skilled-activity capture.
What should a public-dataset comparison include?
Compare domain, viewpoint, modalities, annotations, scale when sourced, access path, license terms, best uses, known gaps, and commercial-use caveats.
What is the difference between public availability and commercial suitability?
Public availability means a portal, download, or gated application exists — nothing more. Commercial suitability depends on the release's license, contributor consent, privacy terms, and whether your intended use is permitted. A dataset can be public and still be noncommercial (EPIC-KITCHENS-100 is CC BY-NC), and a permissive card license does not by itself clear the rights of the people recorded. Treat public access as the start of due diligence, not the end.
Which egocentric datasets allow commercial use?
Verify each release individually — the answer is not uniform. Egocentric-10K's dataset card states Apache-2.0 for the release, but that still leaves consent, privacy, and publicity to review. EPIC-KITCHENS-100 (CC BY-NC 4.0) and EgoDex (CC BY-NC-ND) are noncommercial by their public terms. Ego4D and Ego-Exo4D are governed by signed access agreements that need legal review before commercial training. No public egocentric corpus should be assumed commercially cleared on availability alone.
Looking for best public egocentric video datasets?
Specify modality, task, environment, requested rights posture, and delivery format. Truelabel routes the request to candidate capture partners and helps scope consent/provenance artifacts and commercial licensing requirements for buyer review before delivery.
Compare public dataset limits with a custom collection plan