Markerless Mocap Camera Selection

technology
A source-bounded, camera-agnostic guide to selecting and qualifying cameras for a flexible markerless motion-capture lab.
Author

Dieter Olson

Published

September 27, 2026

Provisional Shop Recommendation

For a shop system intended to observe fast whole-body motion and eventually use six to eight cameras, the current evidence supports a two-camera pilot, not an immediate full-rig purchase. The provisional first candidate is the color Teledyne FLIR Blackfly S BFS-U3-16S2C-CS; the Basler ace 2 a2A1920-160ucBAS is the matched alternate. Both are global-shutter industrial cameras with hardware-trigger support and replaceable lenses. The FLIR candidate trades resolution for a higher vendor-specified maximum rate (226 frames per second at 1440 by 1080 in BayerRG8 with a 10-bit ADC setting); Basler documents 164 frames per second at its default 1920-by-1200 settings. Resolution, pixel format, exposure and firmware must be recorded together: a model name or headline maximum is not a complete acquisition mode.

The distributed eight-camera reference-rig evaluation hypothesis remains the Allied Vision Alvium G5-203 (Alvium G5-203c), with the LUCID Triton2 TRT016S-CC as an Ethernet/PTP challenger. These distributed candidates operate over 5GigE and 2.5GigE Vision with IEEE 1588 PTP synchronization, addressing the cabling and host-controller bottlenecks of high-rate USB topologies across large capture volumes.

The Stereolabs ZED X One GS remains an evaluation candidate for spaces where synchronized GMSL2 cabling is valuable. It is not the default fast-motion choice: its full-resolution mode is 60 frames per second, its 120-frame-per-second mode is binned to 960 by 600, its vendor dual-camera hardware synchronization claim is 15 microseconds, and the complete topology adds capture-card, NVIDIA, fixed-lens, and SDK-version constraints.

This recommendation is deliberately provisional. Camera specifications alone cannot establish pose, joint-angle, or C3D accuracy. A successful two-camera USB pilot does not establish an eight-camera distributed topology, and an attractive camera-body specification does not qualify the complete lab. First define the intended output and its error budget, obtain comparable quotes, and qualify a matched pair. Evaluate distributed Ethernet candidates separately for the full room. A conditional decision to buy two cameras for the pilot requires the lens, host and software review below; no candidate has already earned that decision through a physical test.

WarningNo Procurement Approval & Price Scoping

No listed camera is approved for a full deployment. The earlier camera-body price observations are historical, not current quotes. Current comparable regional prices are unavailable. Complete system costs remain unavailable until formal quotes include lenses, GPIO/trigger hardware, host controllers, topology-specific power/network equipment, cabling, mounting, lighting, storage, spares, taxes, and shipping. The machine-readable registry defaults procurement to deny.

Governed Evidence

The canonical dataset is the camera evidence registry. Its schema is affinedrift/mocap-camera-evidence-registry/v1, and scripts/verify_mocap_camera_registry.py rejects missing sources, review dates expired relative to the registry date, unknown source IDs, cross-camera claim ownership, incomplete purchasing attributes, and status inconsistencies. Those checks establish structural consistency; they cannot establish that a source supports a value. The incorrect camera modes corrected in this review previously passed those checks.

Evidence classes have intentionally different authority:

Evidence Class What It Can Establish What It Cannot Establish
vendor_spec A vendor’s dated statement about its product or SDK Sustained rig performance or independent accuracy
peer_reviewed_observed What a cited study observed under its own protocol Transfer to another camera, task, pose model, or lab
engineering_inference A transparent design screen or pilot decision A qualified threshold or procurement approval
unavailable That the necessary value is not currently authoritative A value to be silently estimated or substituted

AffineDrift owns this public evidence and explanation only. Tools owns camera, time, frame, calibration, and interchange contracts. UpstreamDrift owns capture orchestration, calibration execution, reconstruction, validation, and operator workflows. This page does not implement or duplicate either runtime.

Candidate Matrix

Candidate Shutter Vendor-Specified Video Mode Synchronization Optics Data Path Camera Body List Price Current Role
Teledyne FLIR BFS-U3-16S2C-CS Global 1440 by 1080 at up to 226 fps; BayerRG8, 10-bit ADC Hardware or software trigger CS mount; verify any C-mount adapter USB 3.1 Gen 1 Current regional quote unavailable Preferred fast-motion pilot candidate
Basler a2A1920-160ucBAS Global 1920 by 1200 at 164 fps (default settings) Hardware trigger; programmable API C mount USB 3.0 Quote required (unavailable) Higher-resolution fast-motion alternate
Stereolabs ZED X One GS Global 1920 by 1200 at up to 60 fps; 960 by 600 at up to 120 fps External synchronization; vendor claim within 15 microseconds Fixed narrow or wide GS variant GMSL2 plus capture hardware Current regional quote unavailable Long-cable topology evaluation
Allied Vision Alvium G5-203c Global 1632 by 1248 at up to 225 fps; datasheet specifies Mono8 Hardware trigger; IEEE 1588 PTP; Action Commands C mount 5GigE (PoE) Quote required (unavailable) Distributed reference-rig evaluation hypothesis
LUCID Triton2 TRT016S-CC Global 1440 by 1080 at up to 166.3 fps; mode-dependent Hardware trigger; IEEE 1588 PTP; check action support C mount 2.5GigE (PoE) Quote required (unavailable) Distributed Ethernet/PTP challenger

Primary product evidence is available from:

The original review dates were 2026-08-26 and 2026-08-27; corrected specifications were checked on 2026-09-27. The registry preserves claim-level dates and source limitations. In particular, Basler now documents a 1936-by-1216 full sensor region and a 1920-by-1200 default region; do not combine the larger region with the default-rate claim. The Allied color-model datasheet itself labels its 225-fps condition Mono8; verify Bayer/RGB performance separately. Its RGB8 default does not mean that three-byte color output runs at the one-byte benchmark rate.

Teledyne FLIR Fast-Motion Pilot Candidate

The FLIR candidate leads the two-camera pilot because temporal sampling is the first screen for rapid motion. The vendor lists a global shutter, 226 frames per second at full 1440-by-1080 BayerRG8 output with 10-bit ADC, a CS-mount lens interface, and hardware trigger support. The separate I/O documentation describes the physical input and its propagation specification. None of these statements proves exposure alignment between two assembled cameras, reliable eight-camera streaming, or pose accuracy.

The cited Spinnaker agreement contains camera-use restrictions and separate terms for open-source components. It does not establish the license of a future SDK build or change the license of the open core. A camera adapter must remain optional and satisfy the public Tools camera protocol. Synthetic and prerecorded drivers must continue to work when Spinnaker is absent.

Basler Higher-Resolution Alternate

The Basler candidate offers 2.3 megapixels at 164 frames per second under default settings, global shutter, C-mount optics, a hardware-trigger input, and GenICam-facing pylon APIs. Its additional pixels may improve keypoint localization when the person occupies a smaller portion of the image; that possible benefit is an inference to test, not a result. The lower maximum frame rate may be immaterial for some body segments and insufficient for others.

The pylon SDK is also governed by a vendor license. It belongs behind an optional adapter. A commercial release must review the exact pylon license shipped with the chosen SDK version rather than relying on this summary.

Stereolabs Long-Cable Evaluation

The ZED X One GS offers a synchronized, locking GMSL2 topology and fixed narrow or wide GS lens variants. Stereolabs technical documentation states that dual ZED X One hardware synchronization is accurate to 15 microseconds, and multi-device ZED Box documentation states approximately 15 microseconds. This makes the system interesting where USB cable length and connector reliability dominate. It does not remove the need to measure exposure timing, transport drops, or model latency on the assembled system.

The comparison must preserve mode context. 120 frames per second applies to the 960-by-600 binned mode, while full 1920-by-1200 output is limited to 60 frames per second. A pair is required for ZED SDK stereo or depth features. The SDK and capture topology depend on supported NVIDIA hardware, and their licenses and redistribution terms remain separate review gates.

Allied Vision Distributed Reference-Rig Hypothesis

The Alvium G5-203c remains an Ethernet candidate, with a Sony IMX422 global-shutter sensor, 1632-by-1248 output, 5GBASE-T and PoE. Its model-specific datasheet lists 225 fps at a 525-MByte/s setting in Mono8. Treat that mode condition as part of the specification, including for this color variant. The datasheet lists hardware triggering, action commands and PTP.

Longer Ethernet runs can simplify placement, subject to the selected cable, connectors and vendor link qualification. A 5-Gbit/s interface is a capacity, not the traffic of every mode. Size each shared link, switch egress, host bus and recording path from the actual payload and measured overhead. Synchronized acquisition may create bursts even when average traffic fits.

Vimba X is the vendor SDK. Its licensing information identifies vendor terms and component licenses. Preserve the optional adapter boundary and review the exact deployed components; protocol compatibility alone does not grant redistribution rights.

LUCID Triton2 Distributed Ethernet Challenger

The LUCID Triton2 TRT016S-CC provides an Ethernet/PTP challenger for the distributed reference rig. Its Sony IMX273 global-shutter sensor provides 1440-by-1080 output at a stated maximum of 166.3 fps, with 2.5GBASE-T, PoE and PTP. The vendor states a 10% device-link reserve for resend; record the actual pixel format and reserve in the capture configuration.

The previous 226-fps claim was incorrect: even uncompressed eight-bit pixels would need 2.812 Gbit/s, exceeding 2.5 Gbit/s before overhead. A slower link does not reduce the bytes of a fixed acquisition mode or guarantee lower CPU load. The Arena SDK provides vendor camera APIs and GenICam access; verify the intended integration path and release-specific license rather than assuming a GenTL producer or redistribution permission.

Topology Comparison

Parameter Two-Camera USB Pilot Distributed Ethernet Reference Rig Long-Cable GMSL2 Rig
Representative Models FLIR BFS-U3-16S2C-CS, Basler a2A1920-160ucBAS Allied Vision Alvium G5-203c, LUCID Triton2 TRT016S-CC Stereolabs ZED X One GS
Physical Transport USB 3.0 / USB 3.1 Gen 1 5GBASE-T / 2.5GBASE-T Ethernet (PoE) GMSL2 with FAKRA connectors
Cable Qualification Qualify the exact passive or active cable and controller Qualify cable category, length, connector and link speed Qualify camera, coax cable and capture-card combination
Synchronization Compatible common hardware trigger; measure exposure skew PTP clock alignment plus verified trigger/scheduling support Vendor 15 µs hardware-sync claim for supported topology
Power Distribution USB bus power (host controller dependent) Power-over-Ethernet (802.3af/at switch) Power over Coax / capture card
Host Interface Qualified controller groups and host buses NICs and switches sized for aggregate traffic Supported deserializer/capture hardware and compute
Driver Interface Qualified Spinnaker or pylon adapter Qualified Vimba X or Arena adapter Supported ZED SDK or lower-level driver path
Lab Role Initial fast-motion screening (2 cameras) Full-room distributed rig (8 cameras) Fixed-lens long-cable evaluation

Scoped Pricing vs Total System Cost

Earlier records contained USD 371 for FLIR and USD 399/424 for ZED camera bodies. Those observations are not current offers. On rechecking, the FLIR listing showed a different dollar amount and the old ZED store URL redirected without a matching configuration price. The registry now marks current regional prices unavailable and preserves the historical amounts only in its limitations. Obtain dated, configuration-specific quotes for all five candidates.

A complete mocap lab requires extensive infrastructure beyond the camera body:

  1. Optics: High-resolution C/CS-mount lenses with manual focus/aperture locks with a dated quote for the chosen lens.
  2. Synchronization & I/O: Shielded trigger breakout cables, opto-isolated distribution amplifiers, or a qualified PTP topology; a separate grandmaster is not universally required.
  3. Cabling: High-flex locking USB cables, Cat6A shielded Ethernet cables, or FAKRA GMSL2 cables.
  4. Host & Ingest: Multi-channel PCIe host controllers, multi-Gigabit PoE switches, dedicated NVMe scratch arrays.
  5. Mounting & Structure: Rigid wall/truss mounts, vibration dampeners, protective enclosures.
  6. Illumination: High-CRI flicker-free LED lighting panels.
  7. Spares & Logistics: Cold spares, shipping, regional import tariffs, and taxes.

Complete system cost comparisons remain unavailable until comprehensive, dated bills of materials are quoted for each candidate topology.

Selection Calculations

Temporal Sampling

For locally constant image speed \(v_{px}\) in pixels per second and a camera rate \(f_s\), the nominal inter-frame displacement is

\[ \Delta p = \frac{v_{px}}{f_s}. \tag{1}\]

For varying motion, displacement is the integral of image velocity over the actual frame interval; the expression is a local screen. It is not a Nyquist criterion or an impact-duration measurement. Pixel speed must be evaluated after choosing field of view and resolution. This equation does not choose a universal frame rate. Measure \(v_{px}\) for the intended body, club, implement, or tool region, then set a task-specific maximum displacement. A body-pose model may tolerate a displacement that is unacceptable for a fast implement. The registry’s 120-frame-per-second pilot screen is therefore provisional.

Motion Blur

If exposure time is \(t_e\), the approximate image-space motion blur is

\[ b_{px} = v_{px} t_e. \tag{2}\]

For a declared blur budget \(b_{max}\), the exposure contract is

\[ t_e \leq \frac{b_{max}}{v_{px}}. \tag{3}\]

A high frame rate does not automatically create a short exposure. Lighting, aperture, gain, sensor response, and flicker determine whether the required exposure is usable. A global shutter aligns rows within an image but does not eliminate motion blur or synchronize different cameras. The pilot must record exposure, gain, illumination, and observed blur for every camera.

Raw Payload Screen

An uncompressed payload screen for width \(W\), height \(H\), frame rate \(f_s\), and delivered bit depth \(q\) is

\[ R_{raw} = W H f_s q. \tag{4}\]

Here \(q\) is the delivered bits per pixel, including color channels and packing, not the ADC setting. Bayer8 uses 8 bits (one byte) per pixel; RGB8 uses 24 bits (three bytes). A 12-bit sample stored in a 16-bit container consumes two bytes; packed 12-bit output uses 1.5 bytes before row padding. Debayering on the host increases downstream memory traffic even when the camera link carries Bayer8. Use the driver’s actual payload size when chunks, padding or compression are present.

Arithmetic Screen Pixels and Rate One Stream (Gbit/s) Eight Streams (Gbit/s)
FLIR BayerRG8, 10-bit ADC 1440 × 1080 at 226 fps 2.812 22.494
Basler eight-bit default-region screen 1920 × 1200 at 164 fps 3.023 24.183
Allied datasheet Mono8 condition 1632 × 1248 at 225 fps 3.666 29.329
LUCID eight-bit maximum-rate screen 1440 × 1080 at 166.3 fps 2.069 16.552

These are decimal payload rates, excluding transport overhead, resend, file headers and copies. They are not measured sustained modes. In particular, confirm the Basler and LUCID format/rate combination at the driver, and do not transfer Allied’s Mono8 rate to RGB8. FLIR’s own table reduces full-resolution BayerRG8 from 226 to 165 fps when its ADC changes from 10 to 12 bits. An eight-camera Allied screen needs about 3.666 GB/s of image storage before overhead; interface capacity, PCIe sharing, buffer bursts and sustained disk writes all require separate checks.

Ethernet is one possible architecture, not a requirement imposed by camera count. Independent USB controller groups or multiple acquisition hosts are alternatives if the complete timing and throughput contract passes. Conversely, one nominal 10GbE uplink cannot carry the eight-camera payloads above uncompressed at these rates. A topology diagram must identify every shared bottleneck.

The acquisition contract must measure successful payload, dropped and duplicate frames, buffer occupancy, host-controller/NIC assignment, CPU load, storage write rate, and graceful recovery. A nominal interface bandwidth label is not a per-camera end-to-end guarantee.

Clock Alignment, Exposure Timing and Reconstruction

PTP aligns clocks. A common hardware trigger or supported scheduled action determines when acquisition begins; trigger latency and exposure duration then determine the image’s effective time. Packet arrival time is a different quantity. LUCID’s synchronization example configures PTP and trigger actions separately, and schedules transfer to manage bursts. That example uses other models; confirm equivalent nodes and firmware on the candidate TRT016S-CC.

For uniform integration during locally constant motion, use the exposure midpoint as a first approximation to image time. Two equal start times with unequal exposures have unequal midpoints. At image speed \(v_{px}\) and midpoint mismatch \(\delta t\), the correspondence displacement is approximately \(|v_{px}\delta t|\). Clock offset, clock drift, trigger jitter and exposure timing must therefore be measured separately under load. A timing figure is not itself a 3D error bound.

For an explicitly illustrative \(10{,}000\) pixels/s point, 200 fps gives 50 pixels between frames; 100 microseconds exposure gives approximately one pixel of blur; 50 microseconds midpoint skew gives approximately half a pixel of cross-view mismatch. Increasing frame rate fixes neither of the latter two quantities automatically. A 40 m/s point travels 0.6 mm in 15 microseconds, but triangulation may amplify that inconsistency and motion direction matters. These are constructed examples, not measurements of a golfer or any camera.

Triangulation Geometry

For calibrated, rectified stereo with parallel optical axes, focal length \(f\) in pixels, baseline \(B\) and positive horizontal disparity \(d\), the ideal pinhole relation is

\[ z=\frac{fB}{d}, \qquad \sigma_z\approx\frac{z^2}{fB}\sigma_d. \tag{5}\]

This is a first-order standard-deviation estimate with fixed calibration and small disparity error. If left and right localization errors \(e_L,e_R\) have standard deviations \(\sigma_L,\sigma_R\), then \(\sigma_d^2=\sigma_L^2+\sigma_R^2-2\operatorname{Cov}(e_L,e_R)\). Equal independent errors give \(\sigma_d=\sqrt{2}\sigma_{px}\); correlated errors do not. With \(f=1000\) pixels, \(B=1\) m, \(z=4\) m and independent 0.5-pixel errors, the estimate is 11.3 mm. A positive ten-pixel disparity bias instead changes the reconstructed depth from 4 m to approximately 3.846 m. Random localization precision and systematic timing bias are distinct.

For arbitrary camera placement, linearize the actual calibrated projections. If \(J\) maps a 3D point perturbation to stacked image-coordinate perturbations and image covariance \(\Sigma_u\) is positive definite, a locally identifiable weighted least-squares point estimate has

\[ \Sigma_X\approx\left(J^\mathsf{T}\Sigma_u^{-1}J\right)^{-1}. \tag{6}\]

This conditional estimate requires full column rank and a valid local noise model. Include calibration uncertainty and timing parameters in a joint model or propagate them separately; the formula alone omits them. Near-parallel rays make depth poorly conditioned. Wider baselines can improve geometry while increasing occlusion and appearance differences. More cameras help only when their observations add usable information; repeated biased detections are not independent evidence. Detector confidence is not automatically a calibrated variance.

From Image Error to Golf-Swing Inference

Camera selection affects the argument we can make about a swing. Exposure blur changes landmark localization; timing mismatch combines different poses; calibration and view geometry transform those image errors into anisotropic 3D errors. Fitting a skeleton adds joint definitions, anthropometry and constraints. Filtering trades noise against temporal detail and may shift estimated events. A smooth reconstruction can therefore hide uncertainty rather than remove it.

Differentiation is especially sensitive. For independent zero-mean position noise with standard deviation \(\sigma_X\) in one coordinate, sampled every \(\Delta t\), the unfiltered central-difference velocity and acceleration noise are

\[ \sigma_v=\frac{\sigma_X}{\sqrt{2}\Delta t}, \qquad \sigma_a=\frac{\sqrt{6}\sigma_X}{\Delta t^2}. \tag{7}\]

At 200 Hz and \(\sigma_X=1\) mm, these are about 0.141 m/s and 98.0 m/s². These values describe those stencils and an independent-noise model, not a calibrated mocap system. Correlated noise and filtering change the result. They explain why a position error acceptable for visualization may be unacceptable for acceleration or a brief sequencing claim.

Finally, inverse dynamics needs acceleration, segment inertial properties and external loads in addition to pose. Camera data alone do not uniquely measure muscle forces or identify neural commands. State which load measurements, contact assumptions and regularization supply missing information. Validate positions, joint definitions, derivatives and force estimates at their own level before using them to argue for a mechanism of the golf swing.

What the Research Establishes

Nakano and colleagues evaluated five synchronized cameras at 1920 by 1080 and 120 hertz and reported task- and keypoint-dependent markerless error. Its two participants performed walking, jumping and throwing; it also tested 3840-by-2160 video at 30 Hz. Resolution and rate changed together, so this is not an isolated frame-rate experiment. The study does not qualify any industrial camera or golf-swing endpoint here.

OpenCap used 720-by-1280 smartphone videos at 60 hertz and synchronized keypoints in processing using cross-correlations of their velocities. This is software temporal alignment, not evidence of hardware-synchronized exposure. It is an important accessible baseline, not evidence that 60 hertz is sufficient for every shop task or that its accuracy transfers to a different camera, detector, skeleton, or calibration pipeline.

Song and colleagues compared markerless and marker-based estimates across eight movements and reported larger hip differences, especially for rapid movement such as running, in ten participants. That abstract-level observation does not establish a universal speed-error law. Agreement with marker-based estimates also is not proof of absolute anatomical accuracy. This reinforces a central rule: qualification is task-, joint-, model-, and protocol-specific.

Two-Camera Pilot Protocol

  1. Obtain complete quotes for two FLIR cameras and two comparable Basler cameras. Include matched lenses, locking data cables, isolated trigger distribution, mounts, tripods or structure, lights, host controllers, storage, taxes, shipping, and one spare strategy.
  2. Review the exact SDK versions and EULAs. The MIT core must operate without either vendor SDK.
  3. Select lenses from the measured capture-volume dimensions. Record field of view, working distance, focus, aperture, and distortion; do not select only by nominal focal length.
  4. Bench two matched cameras at the intended resolution, bit depth, frame rate, exposure, and cable length. Measure exposure skew, drift, frame loss, duplicates, jitter, and recovery.
  5. Calibrate a real two-camera geometry, retain every observation and solver version, and report residuals, covariance, ray-angle coverage, and movement invalidation.
  6. Run the intended pose backend on representative clothing, lighting, occlusion, body sizes, and motions. Keep detector confidence separate from geometric uncertainty.
  7. Compare triangulated points and derived joint quantities against a declared reference. Report all unavailable and failed keypoints; never interpolate them silently.
  8. Repeat at the worst intended capture-volume locations. If the intended output includes club motion, derivatives or dynamics, validate those outputs and their required load/model inputs as well as body pose. Only then compare rate and resolution against the declared task budget.
  9. Add cameras in stages. Re-run synchronization, bandwidth, calibration, visibility, thermal, storage, and pose tests before scaling to an eight-camera distributed topology.

Required Qualification Evidence

The full-rig decision remains blocked until the following evidence exists:

  • a dated bill of materials and exact license review;
  • a camera capability record from the real driver rather than a product-name switch;
  • sustained stream results at every supported camera count and mixed-capability case;
  • exposure timing for the chosen trigger/clock architecture, plus drift, drop, duplicate, and recovery measurements;
  • intrinsic and extrinsic calibration observations, residuals, covariance, and validity state;
  • task-volume visibility and ray-angle maps with occlusion and lighting challenges;
  • reference comparisons for the actual pose model, skeleton, filtering, and biomechanical outputs;
  • C3D export and independent-reader round trips with units, rates, residuals, events, and explicit loss reporting;
  • privacy, consent, retention, security, SBOM, and redistribution approval; and
  • an exact-version commissioning report that states what the rig cannot measure.

Extension Contract

A new camera does not require a core code change when it satisfies the shared provider protocol. Its adapter must report capabilities such as stream modes, shutter, trigger, clocks, controls, pixel formats, lens metadata, and cancellation. Unsupported functions return typed unsupported outcomes. The lab manifest selects providers and records immutable device identities and settings; the reconstruction code consumes canonical observations and never switches behavior by vendor name.

To add a camera to this public registry, add its dated primary sources, exactly one claim for every purchasing attribute, limitations, review dates, and an explicit adapter state. If price, sync, license, or topology is unknown, record unavailable. The verifier will reject an apparently complete camera assembled from missing or cross-model claims.

Current Limitations

  • No candidate has been connected to this workstation under the canonical Tools protocol.
  • No physical timing, blur, lighting, field-of-view, pose, triangulation, or C3D test has run.
  • Complete system costs remain unavailable until comparable complete regional quotes exist.
  • Vendor SDK summaries are not legal opinions.
  • The cited studies bound only their own cameras, participants, models, tasks, and protocols.
  • This document does not approve human-subject recording, public release, or biomechanical use.

Begin with a declared measurement target and complete quotes. Select a matched pilot pair only after its mode, optics, timing, software and host requirements are supportable; expand after the measured results satisfy that target. No specification table can substitute for this chain.

Citation

BibTeX citation: Download .bib
@online{olson2026,
  author = {Olson, Dieter},
  title = {Markerless {Mocap} {Camera} {Selection}},
  date = {2026-09-27},
  url = {https://affinedrift.com/articles/markerless-mocap-camera-selection.html},
  langid = {en}
}
For attribution, please cite this work as:
Olson, Dieter. 2026. “Markerless Mocap Camera Selection.” September 27. https://affinedrift.com/articles/markerless-mocap-camera-selection.html.