L

Lumens Pose Observatory

MediaPipe C++ · tiles, tracking, and real classification

Checking engine…

Analysis

Select video No file selected

PoseLibrary

Max people Target tile Resync (frames) Rotation skip Refresh budget Detector conf.

CLAHE (Lab, L channel)

CLAHE clip limit CLAHE grid size

Gamma correction

Gamma value

Track rescue (experiment) Pose fallback ML (experiment)

Only affects tracks the C++ rules leave as UNKNOWN. Measured against ground truth: HistGradientBoosting now improves end-to-end accuracy over no fallback (0.841 vs 0.833 aggregate); Logistic Regression is close but still slightly below (0.817) - see lumens-benchmark/dataset_prep/. Applies on the next analysis.

Advanced parameters

Pose confirmation Max misses Dedup IoU Containment Threads

Diagnostic layers

Tiles discovery Refinements ROI rotation Raw detections Bounding boxes Skeleton/landmarks Labels & IDs Show preprocessed (CLAHE/gamma)

People0

Standing0

Sitting0

Hand raised0

Total frame0 ms

Phaseidle

Waiting for video0%

MediaPipe ViewerSelect a video to inspect the real tiles, refinement, and rotation calls.

Lumens Pose Observatory

Processed video

Active tracks

Frame Detect() calls