Paired perspective / 03

Two Viewpoints Become Useful When They Share a Reliable Timeline.

Synchronized ego/exo data connects what the actor could see with what the environment reveals—supporting cross-view correspondence, occlusion recovery, and learning across embodiments.

EGO

Actor-visible intent, hands, objects, and local state.

EXO

Whole-body motion, geometry, context, and hidden events.

Paired Capture Is a Systems Problem

Concurrent recording is not enough. The dataset needs evidence that streams correspond in time, space, task, and identity.

Synchronize

Declare the clock source, timestamps, trigger method, acceptable drift, and long-take verification procedure.

Calibrate

Store camera intrinsics, extrinsics, coordinate frames, recalibration events, and known spatial error.

Associate

Map takes, actors, objects, tasks, and annotations across views without exposing personal identity.

Use cases

Pay the Pairing Cost Only When the Model Needs Correspondence.

Viewpoint-invariant representation learning
Human-to-robot demonstration transfer
3D pose, hand pose, and object trajectory recovery
Occlusion-aware action understanding
Cross-view retrieval and language grounding
Independent evaluation of actor-visible decisions

Acceptance Criteria for a Paired Pilot

Every criterion should be measurable before collection expands.

Time

Alignment

Maximum drift, missing frames, clock resets, and verification points.

Space

Calibration

Coordinate conventions, error tolerance, scene changes, and recalibration.

Coverage

Visibility

Critical actions visible in one or both views with documented occlusion limits.

Structure

Association

Stable take, stream, task, and annotation keys across the synchronized release.

Governance

Rights

Each view cleared for its intended use and distribution tier.

Delivery

Ingest

Buyer loader proves aligned decoding, timestamps, and episode boundaries.

Paired pilot

Define the Correspondence Your Model Must Learn.

Tell us the views, task, time tolerance, calibration needs, modalities, and target representation.

Scope Paired Capture