Evidence-backed summary for YouTube video iS3G8p90TYY. Built from the verified 79-minute 1080p recording, dense 30fps checks, YOLO person/ball candidates, pose hints, annotated clips, and timestamped event tables.
Scrub the video and inspect the same timestamp as player tracks, pose-derived actions, the ball hypothesis, optical-flow evidence, contact candidates, and physics-fit diagnostics. The 3D scene is draggable.
8 event windows10 fps states2.5D normalized, not metric
Evidence video
loading…Loading manifest…
H06 has no model ball anchors; its candidate path is hidden by default but remains inspectable through the low-confidence toggle and JSON.
Constrained above-court orbit: drag to rotate; mouse wheel to zoom. Height/depth are inferred proxies with visible uncertainty—not metric measurements.
ObservedSource pixels, fixed scene, net and sand landmarks.
DetectedPose model, sports-ball model, motion and optical flow.
InferredTrack interpolation, contact/crossing, normalized 2.5D and ballistic fits.
UnknownMetric depth, height, speed, exact score and point winner.
Bottom line
The analysis supports highlight/action identification and player-involvement observations. It does not support exact scoring or a winner without inventing.
Best highlight:01:12:38–01:12:45, late net action / possible defensive dive.
Dedicated MOG2/KNN/background-difference pass over the 8 highlight windows. This layer isolates movement timing and movement regions; it supports activity confidence, but it still does not prove touch order, point winner, or score.
Generated locally from verified artifacts in /home/kristjan-variksoo/volleyball-analysis-iS3G8p90TYY. All sports claims are timestamped; unresolved score/winner claims are intentionally not invented.