Motion
EVIDENCE STRAND // 014SIGNAL STATE: VERIFIEDGPU RESULT: NOT YET MEASURED

Measured output. Named limits.

Two pose engines processed the same deterministic 288-frame sample on one CPU system. This record publishes the aggregate measurements, method, consistency result and unresolved limits without redistributing the private source video or raw result JSON.

Keypoint R-CNN527.840steady-state seconds
MediaPipe26.925steady-state seconds
Runtime ratio19.60×not a quality comparison
Model forward98.99%of Keypoint R-CNN total

One identity. Two engines.

The configurations reject a different input hash, frame count or sampled-frame identity. Setup and model initialisation are timed separately from the published steady-state totals.

Input
Private consented tennis video
Source
863 frames · 29.948 FPS · 28.817 s
Sample
288 frames at a 10 FPS target
Compute
CPU · FP32 · batch 1
Processor
Intel Core Ultra 7 258V · 8 cores
Memory
31.51 GB available
Software
Python 3.11.9 · PyTorch 2.13.0+cpu
Provenance
Benchmark commit ccec398e5c38

The private footage and identifiable outputs are excluded. The downloadable file contains de-identified aggregate evidence only.

The model forward pass is the bottleneck.

Keypoint R-CNN527.840297 s · 0.546 sampled FPS
MediaPipe26.925353 s · 10.696 sampled FPS
Keypoint R-CNN stageSecondsShare of total
Decode + sampling4.1032280.78%
Inference envelope523.41685599.16%
Model forward522.49889098.99%
Person selection0.009820<0.01%
Angle analysis0.011832<0.01%
Serialization0.2454710.05%

The inference envelope includes model forward and surrounding inference work. Model-forward time is shown separately and should not be added to the envelope.

Repeatable output, not claimed accuracy.

Two retained Keypoint R-CNN CPU repetitions were compared on eight common frames. The result validates same-model output consistency under the tested conditions.

Detection presence1.0
Detection count1.0
Selected person1.0
Joint-angle MAE0.0°
Evidence boundary

This is not ground-truth pose accuracy, biomechanics validity or coaching accuracy. Cross-model figures describe agreement between model families only.

Targets, not results.

Planned

The audited machine has no CUDA device. GPU, reduced-precision, TensorRT and Triton performance remain unmeasured.

End-to-end throughput0.546 sampled FPS · CPUAt least 2.73 sampled FPS · 5×
Inference throughput0.550 sampled FPS · CPUAt least 5.50 sampled FPS · 10×
Edit-run-profile cycle12.779467 s · CPU medianAt most 10 s after setup
Output validationCPU repeat consistency passesCUDA FP32 and reduced precision within declared thresholds

The open questions stay visible.

  • Ground-truth pose or biomechanics accuracy.
  • CUDA FP32 and reduced-precision performance.
  • Behaviour under occlusion, low light and multiple-athlete scenes.
  • Validated sport-specific coaching interpretation.
  • Metric distance without camera calibration or a court scale.

Public aggregate generated from the internal benchmark package. Raw media and identifiable frame outputs are intentionally absent.

Return to Motion