Human motion-capture datasets: pose, trajectories, video alignment and licensing

Motion capture is more specific than a video of movement

A human-motion source can contain optical markers, inertial measurements, estimated joint positions, a rigged body representation, synchronized cameras or some combination. The target task determines which are necessary. Animation may need retargetable motion; pose estimation may need image/motion pairing; motion generation may need a consistent representation and action context. A collection of ordinary video can be valuable without containing measured three-dimensional trajectories. Preserve that distinction in both sourcing and public copy.

Human motion should also remain separate from robot-action data. A joint trajectory for a person does not supply robot torque, actuator commands or a compatible control policy. Retargeting is a transformation with assumptions, not a change of label. If a buyer needs object contacts, forces, task outcomes or interaction geometry, those fields require their own evidence. Do not imply they exist because a motion file plays convincingly on a character.

Public benchmarks and source-specific conditions

AMASS describes a common body-model representation built from multiple marker-based source datasets; its original release reports more than forty hours and over eleven thousand motions. These are fitted model parameters, not the same thing as untouched marker measurements. Its official license restricts the standard release to noncommercial purposes and explicitly addresses commercial training, while inviting separate commercial licensing inquiries.

The CMU motion library states that its 2,605 trials are free for all uses and notes important data limitations, including placeholder finger/thumb joints and people who may appear under multiple subject identifiers. Human3.6M’s official description is another pose/image reference; direct access redirected during this review, so current access should be checked. Compare each release’s representation, terms and actual source evidence rather than assuming every motion corpus has identical rights or precision.

Buyers and practical commercial source routes

Potential users include animation tools, avatar systems, motion-model researchers, pose-estimation developers and interaction-analysis products. These are application segments, not verified Fiund customers. A model that predicts future motion may prioritize complete temporal sequences and action transitions; an animation product may prioritize clean retargeting and foot contact; a pose benchmark may need calibrated image correspondence. Define the deliverable around the use, not the broad term “mocap.”

Sources can include original capture studios, authorized production archives, commissioned performer sessions and research releases under their own conditions. AMASS’s commercial contact is a real route to discuss its own terms, not a universal license to all human-motion sources. A studio’s standard performance or animation license may not cover model training or source redistribution. Review performers, source authority, body-model software and intended uses as distinct parts of a proposal.

Skeletons, coordinates and units must be explicit

A motion file is difficult to use reliably without a skeleton definition. Record joint names, hierarchy, rest pose, bone-length conventions, rotation representation, units, up axis, handedness and root motion. A position can be global, root-relative, camera-relative or normalized; these are not interchangeable. A model trained on a mixture without a mapping can learn artifacts or produce physically implausible output even when each file opens successfully.

Rotation and time conventions also matter. Euler angles, quaternions and axis-angle parameters require defined ordering and interpretation. Frame rate should refer to the actual sequence after resampling, with original timing preserved. If motion is smoothed, gap-filled, retargeted or fitted to a body model, record the transformation and version. A clean rendered preview can hide lost detail, foot sliding or an incorrectly interpreted coordinate frame. Review numeric data and visualizations together.

A useful record schema

The manifest should link original performance, subject/session group, capture method, representation, transformations, paired media and use-documentation references. Subject IDs should support the agreed independence analysis without exposing unnecessary personal information. If the source cannot reliably establish subject grouping, record that limitation. Multiple names or session IDs should not be treated automatically as different performers.

Paired video, contacts and outcome labels

If video accompanies motion, define whether cameras were calibrated and how clocks map to the motion sequence. A roughly synchronized reference movie may be adequate for browsing but insufficient for frame-accurate pose learning. If camera parameters or exact timing are unavailable, state that boundary. A task that needs image-to-pose truth should not inherit a stronger claim from an animation-oriented archive.

Contact and interaction labels need their own policy. Foot contact can be annotated, inferred from motion, measured or approximated for animation; those meanings differ. An object interaction may require object trajectories and scene geometry absent from body motion alone. Keep action names, task instructions and success labels separate. A performer can imitate a task without a physical object, which may suit animation but not a manipulation-understanding model. Record performed, mimed and simulated contexts explicitly.

Pricing and a worked project model

No representative public average for commercial human-motion AI datasets is established here. Public releases have different access conditions, including free-use statements and restricted research licenses. Custom prices can include performers, studio/equipment, choreography, capture, cleanup, retargeting, paired video, annotation and permitted uses. A performance minute, motion clip, cleaned take and body-model sequence are different quote units.

For a hypothetical production scenario, suppose a session captures sixty distinct takes and a buyer expects forty-five to pass a defined technical review. If each accepted take requires an assumed eight minutes of cleanup and four minutes of independent check, the modeled work is nine hours. At a chosen loaded rate of $45 per hour, that component is $405. It excludes performer/session costs, source rights, equipment, specialist retargeting, body-model licensing, corrections and delivery. These assumptions are not observed market rates or a Fiund quote.

The accepted yield is important. A quote for raw takes may look cheaper than one for cleaned motion while excluding the work that makes it usable. Compare native and delivered representations, independent performers, task coverage, contacts, paired video and quality thresholds. Ask whether placeholder joints, reconstructed gaps and failed takes count as delivered signal. A research database’s free access should not be presented as the price of a newly commissioned commercial performance collection.

Sample acceptance and source comparison

One archive may contain broad action variety in a legacy skeleton; another may contain fewer performances with consistent calibrated video and a modern body representation. The best choice depends on the task. Normalize the comparison to required signals and use scope rather than selecting the most convenient preview format. Request a sample with numeric files, skeleton definition, representative renderings and any paired source media or labels.

Review coordinate orientation, units, root movement, joint continuity, occlusion gaps, foot sliding, contact consistency and retargeting artifacts. Include fast movement, turns, transitions and the difficult actions the model must handle. A sequence that looks plausible from one camera may reveal errors from another. Check whether body proportions, smoothing or fitting alter the relevant task signal. Separate numeric validity, motion plausibility, capture evidence and permission review.

A concrete retargeting and joint-evidence check

Imagine one performance appears in its native skeleton, a body-model fit and a retargeted game rig. Those are three representations of one motion, not three independent performances. Keep the common source link and document changes to scale, root motion, joint limits and timing. A retargeted hand can look plausible while containing no measured finger data, so appearance alone should not upgrade the evidence claim.

A pilot can inspect numeric trajectories and renderings from several views, including transitions and contacts. Ask which joints are measured, reconstructed, fitted or placeholders and what uncertainty remains. Compare raw versus cleaned versions under the task criteria. A quote can then identify the original source, representation work and permitted uses separately, rather than selling a convenient file format as a universal motion ground truth.

Multi-person interaction needs more than two motion files

Two independently captured performances do not automatically describe a synchronized interaction. If the task needs a handoff, shared object, collision avoidance or coordinated movement, request the common clock, spatial relationship and object context. Retargeting each performer separately can alter relative contact or spacing. A pilot should inspect the combined scene and document which relationships are measured, reconstructed or staged. Keep those requirements separate from ordinary single-person animation cleanup.

Evaluation, deals and Fiund sourcing

Hold related performers, sessions or source datasets out according to the intended generalization. Random windows from one long performance can share nearly identical motion across train and evaluation. Retargeted versions of one performance should remain linked, as should resampled or mirrored derivatives. If subject grouping is uncertain, disclose the limitation instead of counting IDs as independent people. Evaluation on an animation representation should not be described as proof of clinical biomechanics or robot-control suitability.

No verified human-mocap-specific AI-training license transaction amount is established here. Animation asset packs, body-model subscriptions, film production budgets and studio acquisitions are not interchangeable data-license prices. Send Fiund the task, capture method, skeleton/representation, timing, paired media, contacts, performer/session coverage and intended uses. Fiund can assess potential source fit and review needs, with availability, rights, transformations, acceptance and commercial terms confirmed for selected material.

Questions buyers ask