Brief us

Datasets built to order, from the hard parts of the public web.

Text-rich images, egocentric video, multilingual speech and human motion — sourced, annotated and evaluated for your training brief. The datasets below are our portfolio; yours is the product.

Brief us on a build Explore Textphotos

custom builds · private evals under NDA · provenance-verified

Sample viewer

Live annotated samples load here. The full viewers live on the dataset pages.

Explore multilingual images, or switch to our annotated video collections. Open the full sphere →

5.2M+ video IDs screened for eval leakage across 15 public corpora
≈1 in 8 of cut clips survive final curation
10 languages in the talking-head line
100% of records carry a provenance manifest
Our work

On the shelf, built end to end

Explore real annotated samples, dataset specifications and the available quality checks for each collection. Use them to shape the scope of your own build.

image Live demo

Textphotos — Text-Rich Images

Text across languages, scenes and surfaces — explore the images and inspect the words within.

2,000
images in the interactive sample
6
confirmed primary languages
12
scene categories for dataset selection
10,625
detected text regions in this sample
Rights scoped per delivery
video Live demo

Egocentric How-To & Manipulation Video

First-person how-to video, curated for hands-on manipulation density and annotated for action understanding.

3,479
videos screened to cut 3,892 procedure clips — 500 balanced into the reference batch
857 GB
of curated first-person footage on disk, from which the reference batch is drawn
≤4 / ≤6
hard caps per source video and per channel — diversity is enforced, not hoped for
1.0
viewpoint accuracy on the delivered set (25/25), by the annotation stack that ships
Mixed, documented per record
video+audio Live demo

Multilingual Talking-Head

Verified on-camera speech across 10 languages — face tracks, 478-point landmarks, phoneme-to-viseme timelines, emotion.

3.21M
video IDs from 9 public speech corpora screened for overlap — none found in this delivery
2,205
candidates rejected of 2,515 screened, each with a named gate on the record
10
languages with gate-passing content in the current delivery — incl. th, vi, id, ar, ko
478
facial landmark points per tracked frame
Documented per record, person-screened
video Live demo

Human Motion & Pose

Culturally diverse human movement — dance, martial arts, sport — with whole-body 2D pose, hands, face and a 3D body reconstruction on every frame.

2,137 → 8
discovery rows filtered to the clips in this demo — 106 licence-verified byte-for-byte, 76 past the resolution gate
133
2D whole-body keypoints per person — body, feet, face, both hands — plus a full 3D body on every frame
6
culture regions in the demo set — street dance to Chinese martial arts to Okinawan drum dance
98
acceptance checks on the shipped package — 74 at build, 24 by an independent verifier, zero failures
CC BY 3.0 (demo set), documented per record
Custom build

Your dataset

A modality, domain or language mix that isn't on the shelf — built to your spec with the same curation funnel and QA.

Brief us →
Custom data

Custom is the product

The interesting datasets rarely exist yet. Brief us on modality, domains, languages and volume — we build to spec with the same funnel, annotation stack and QA you see in our own lines.

  1. 01
    Brief

    One email: modality, domains, languages, volume, annotation depth, deadline.

  2. 02
    Sample in ~48 hours

    A free eval slice with full annotation — run it through your pipeline and measure.

  3. 03
    Pilot in 1–2 weeks

    Staged delivery with an accuracy SLA verified on an agreed golden set.

How it's made

A funnel, not a firehose

Cheap filters first, expensive ones last: most candidates are rejected from metadata, thumbnails and embeddings alone, before a single full video is downloaded. Pass rates below are from a recent production run.

  1. S0 79% Metadata gates pass rate on duration, language and live-status checks — the cheapest filters run first
  2. S1 82% Frame & embedding probes storyboard frames + viewpoint/domain embedding classification, still with no full download
  3. S2 33% Shortlist for full download only top-scoring candidates are ever fetched at full resolution — 732 of the 3,479 videos discovered
  4. S3 ×5 Lossless cutting each downloaded video yields multiple candidate procedure clips
  5. S4 ≈13% Final curation of cut clips survive balancing by domain, viewpoint, resolution and leakage screening

Full methodology — QA gates, golden sets, leakage screening — documented here.

Provable quality

Measured, not promised

Measured on golden sets during real annotation runs.

1.0 viewpoint accuracy, 25/25 shipping annotation stack on the delivered set · an earlier v1 debug run scored 0.941 on 17 clips
9.64 action segments per minute measured on the delivered set: 858 segments, 112 distinct verbs, 2.9% vague · egocentric line
2,205 candidates rejected, each with a named gate of 2,515 screened · geometry 1,726, sync 1,469, speaking 1,216, resolution 1,063 · talking-head line
2.4% leakage found and removed overlap with COIN, CrossTask, YouCook2, HowTo100M, HTM-AA and VidChapters IDs · egocentric line
Provenance & rights

Every record can explain itself

Training data is an asset only while its rights story holds. We collect from the open public web and document origin down to the individual record.

On every record
  • source URL
  • exact license
  • fetch timestamp
  • transform hash
With every delivery
  • dataset card / datasheet
  • QA report with golden-set metrics
  • SHA-256 manifest, verified at acceptance

Due-diligence questionnaires answered on request. License filtering to your requirements is part of scoping.

Talk to the people who build the pipelines

You reach the engineers directly: straight answers about coverage, formats, licensing and feasibility, and a private eval if it fits.

Brief us Email us