Aug 21 - 6

🔎 Signal Quality ⓘHow much to trust this clip
Each domain is graded by how much evidence the AI actually had — players tracked, ball flight seen, court calibration, shots and reactions detected. Low-quality domains are automatically softened or hidden elsewhere in this report, so numbers you do see are the ones worth reading.

How much to trust each part of this specific clip — graded by the evidence available, not a generic assumption.

🦴 Body Mapping & Posture Analysis ⓘHow it's analyzed
The pose model finds 17 anatomical keypoints (shoulders, elbows, wrists, hips, knees, ankles) on every sampled frame. Those joints are smoothed over time and checked against the player's own body proportions before anything is measured from them, so a joint flung onto the wrong body is dropped rather than averaged in. Angles and reach are measured against each player's torso, which makes a far-court player comparable to a near-court one. Where the player's shot times are known, the same samples are also split into ready / backswing / contact / recovery instead of averaged flat across the clip.

P1

P1

TEAM A

Near left

P2

P2

TEAM A

Near right

P3

P3

TEAM B

Far left at start ⇄

P4

P4

TEAM B

Far right at start ⇄

×

Skeleton overlay — YOLO11x-Pose 17-keypoint estimation

Player 1 · Near left

37 shots detected

Needs Improvement

Knee Angle (L)

162°

Knee Angle (R)

160°

Knee Position

Good

Stance Width

Too Wide

Back Alignment

Good

Weight Balance

Balanced

⚠ Issues (1)

• Stance too wide

💡 Tips to Improve

→ Narrow stance slightly for quicker lateral movement

🏓 Shot Breakdown

Dinks: 5 Drops: 4 Drives: 16 Serves: 5 Returns: 7

📋 Activity Log (37 events) ▼

#	Time	Shot	Velocity
1	0:24	Serves	0.036
2	0:25	Returns	0.036
3	0:27	Drives	0.050
4	0:33	Dinks	0.038
5	0:53	Returns	0.041
6	1:00	Serves	0.076
7	1:01	Returns	0.047
8	1:02	Drives	0.050
9	1:05	Drives	0.078
10	1:08	Drives	0.058
11	1:29	Drives	0.043
12	2:10	Drives	0.040
13	2:12	Dinks	0.048
14	2:15	Dinks	0.051
15	2:30	Serves	0.097
16	2:31	Returns	0.030
17	3:11	Returns	0.111
18	3:27	Drives	0.073
19	4:19	Drops	0.032
20	4:21	Drives	0.069
21	4:22	Dinks	0.035
22	4:23	Dinks	0.025
23	4:29	Serves	0.119
24	4:30	Returns	0.044
25	5:42	Returns	0.030
26	5:46	Drives	0.038
27	5:54	Drives	0.130
28	5:54	Drives	0.038
29	6:10	Serves	0.195
30	6:19	Drives	0.045
31	6:35	Drops	0.036
32	6:37	Drives	0.056
33	6:38	Drops	0.035
34	6:41	Drops	0.040
35	6:50	Drives	0.059
36	6:53	Drives	0.048
37	7:12	Drives	0.045

1852 frames analyzed ⓘPer‑player count
Frames where this player was clearly visible and pose‑tracked — not the whole video. It naturally differs between players (some are on‑screen, unoccluded, or near the camera more than others), so unequal numbers here are expected, not an error.

Player 2 · Near right

22 shots detected

Needs Improvement

Knee Angle (L)

160°

Knee Angle (R)

165°

Knee Position

Good

Stance Width

Too Wide

Back Alignment

Good

Weight Balance

Balanced

⚠ Issues (1)

• Stance too wide

💡 Tips to Improve

→ Narrow stance slightly for quicker lateral movement

🏓 Shot Breakdown

Dinks: 7 Drops: 1 Drives: 5 Serves: 2 Returns: 6 Volleys: 1

📋 Activity Log (22 events) ▼

#	Time	Shot	Velocity
1	0:21	Returns	0.047
2	0:54	Drives	0.047
3	1:11	Dinks	0.040
4	1:33	Dinks	0.106
5	1:45	Returns	0.032
6	2:11	Dinks	0.034
7	2:33	Drops	0.024
8	3:24	Serves	0.024
9	4:07	Returns	0.205
10	4:14	Drives	0.075
11	4:22	Dinks	0.035
12	4:23	Dinks	0.038
13	4:24	Dinks	0.065
14	4:26	Volleys	0.093
15	4:33	Drives	0.055
16	4:44	Returns	0.051
17	5:45	Drives	0.071
18	5:51	Returns	0.089
19	5:53	Dinks	0.077
20	6:10	Returns	0.032
21	6:32	Drives	0.030
22	6:50	Serves	0.111

1529 frames analyzed ⓘPer‑player count
Frames where this player was clearly visible and pose‑tracked — not the whole video. It naturally differs between players (some are on‑screen, unoccluded, or near the camera more than others), so unequal numbers here are expected, not an error.

Player 3 · Far left at start

26 shots detected

Good

Knee Angle (L)

161°

Knee Angle (R)

162°

Knee Position

Good

Stance Width

Good

Back Alignment

Good

Weight Balance

Balanced

🏓 Shot Breakdown

Lobs: 2 Dinks: 5 Drives: 10 Serves: 2 Returns: 6 Volleys: 1

📋 Activity Log (26 events) ▼

#	Time	Shot	Velocity
1	0:14	Dinks	0.036
2	0:20	Serves	0.058
3	0:22	Drives	0.033
4	0:34	Dinks	0.031
5	2:09	Returns	0.093
6	2:11	Dinks	0.064
7	2:13	Dinks	0.049
8	2:15	Drives	0.066
9	3:10	Serves	0.034
10	4:17	Returns	0.017
11	4:18	Dinks	0.023
12	4:32	Drives	0.079
13	5:22	Returns	0.033
14	5:43	Lobs	0.137
15	5:44	Drives	0.044
16	5:45	Drives	0.111
17	5:47	Drives	0.081
18	5:51	Drives	0.062
19	5:52	Drives	0.112
20	5:58	Returns	0.043
21	6:31	Returns	0.051
22	6:32	Drives	0.016
23	6:33	Lobs	0.053
24	6:34	Drives	0.069
25	6:50	Returns	0.088
26	6:59	Volleys	0.033

1169 frames analyzed ⓘPer‑player count
Frames where this player was clearly visible and pose‑tracked — not the whole video. It naturally differs between players (some are on‑screen, unoccluded, or near the camera more than others), so unequal numbers here are expected, not an error.

Player 4 · Far right at start

38 shots detected

Good

Knee Angle (L)

165°

Knee Angle (R)

164°

Knee Position

Good

Stance Width

Good

Back Alignment

Good

Weight Balance

Balanced

🏓 Shot Breakdown

Dinks: 9 Drives: 12 Serves: 13 Returns: 3 Volleys: 1

📋 Activity Log (38 events) ▼

#	Time	Shot	Velocity
1	0:31	Serves	0.039
2	0:32	Returns	0.053
3	0:33	Drives	0.020
4	0:47	Dinks	0.070
5	0:52	Serves	0.026
6	0:54	Dinks	0.016
7	1:03	Drives	0.027
8	1:04	Volleys	0.089
9	1:05	Drives	0.092
10	1:06	Drives	0.023
11	1:16	Serves	0.255
12	1:17	Returns	0.116
13	1:44	Serves	0.049
14	2:09	Serves	0.030
15	2:10	Dinks	0.029
16	2:32	Drives	0.056
17	2:42	Drives	0.066
18	3:11	Dinks	0.024
19	3:26	Returns	0.057
20	3:50	Dinks	0.029
21	4:06	Serves	0.030
22	4:17	Serves	0.012
23	4:17	Dinks	0.027
24	4:18	Drives	0.065
25	4:20	Drives	0.051
26	4:21	Dinks	0.017
27	4:23	Dinks	0.034
28	4:31	Drives	0.027
29	4:43	Serves	0.035
30	5:22	Serves	0.017
31	5:41	Serves	0.185
32	5:50	Serves	0.074
33	5:57	Serves	0.037
34	6:30	Serves	0.009
35	6:34	Drives	0.046
36	6:39	Dinks	0.025
37	6:40	Drives	0.102
38	7:07	Drives	0.055

1363 frames analyzed ⓘPer‑player count
Frames where this player was clearly visible and pose‑tracked — not the whole video. It naturally differs between players (some are on‑screen, unoccluded, or near the camera more than others), so unequal numbers here are expected, not an error.

🏋 Biomechanical Analysis ⓘWhat's measured
Joint angles captured at each shot contact moment using YOLO11-Pose keypoints. Covers swing arc width, racket-arm extension, hip rotation range, and weight-transfer consistency. Requires pose keypoints to be present in the AI result.

Per-Player Arm Mechanics

Player 1

Elbow: 151.2° (Extended)

Arm reach: 0.969 (Wide Reach)

Player 2

Elbow: 153° (Extended)

Arm reach: 0.932 (Moderate)

Player 3

Elbow: 145.1° (Extended)

Arm reach: 0.847 (Moderate)

Player 4

Elbow: 142.3° (Active)

Arm reach: 0.818 (Moderate)

🕐 Shot Events Timeline ⓘHow events are detected
Shot events are auto-detected by peaks in left/right wrist velocity (YOLO11 keypoints 9 & 10). Each event is timestamped, assigned to a player slot via ByteTrack, and classified by contact side (forehand/backhand) based on body orientation.

* Wrist speed is a relative shot-power proxy from keypoint motion at the sampled frame rate — it is not a real-world mph and is not comparable to ball speed. At ~2.7 fps, true swing/ball velocity can't be measured (see the Ball Speed panel's confidence).

📍 Player Court Positioning ⓘHow it's computed
Player bounding-box centroids are mapped to a top-down court view using homographic perspective correction (court corners → standard template). Sampled at ~4 fps throughout the match. Darker areas indicate more time spent in that zone.

Player 1

P1

Left Court cx: 0.277

0.468

Lateral Range

2258

Frames Tracked

Player 2

P2

Right Court cx: 0.724

0.599

Lateral Range

1802

Frames Tracked

Player 3

P3

Center Court cx: 0.522

0.473

Lateral Range

2040

Frames Tracked

Player 4

P4

Center Court cx: 0.574

0.348

Lateral Range

2396

Frames Tracked

Court View — Top Down (2D)

Far baseline

Net

Near baseline

P1

P2

P3

P4

Left sidelineRight sideline

🎬 Scene Analysis ⓘWhat's detected
Video divided into time segments by whether players are present in frame (not rally intensity — that's the motion-based Active Play Rate above). Detects indoor/outdoor environment, camera angle quality, and crowd presence. Useful for filtering out warm-up or dead-time footage before reanalysis.

Detected video segments by player presence in frame

👥 Players Detected ⓘHow detection works
YOLO11 object detector tracks distinct player bounding boxes using ByteTrack multi-object tracking. Consensus voting over a sliding window resolves the final player count. Use the override control below if the AI count is wrong — it affects all per-player analytics.

Player 1

Team A  90% conf.

Player 2

Team A  92% conf.

Player 3

Team B  81% conf.

Player 4

Team B  80% conf.

🏓 Shot Analysis ⓘHow shots are classified
Each detected shot is classified by type (serve, dink, drive, return, volley, drop, lob) using a combination of spatial position on court, ball trajectory angle, and wrist velocity at contact. Counts are aggregated across all players. Confidence is higher with good camera angle and lighting.

🎯 3D Shot Viewer ⓘWhat is measured and what is drawn
Measured: where each shot was struck from (court position at the player's feet), who hit it, the shot type, and the ball dots when the track is verified.
Landings: when the ball track is verified, each landing is the first detected bounce after contact and its arc is drawn SOLID. Otherwise the landing is a stylized estimate and the arc is DASHED. The banner above the court says how many of each.
Always illustrated: arc height — a ground-plane ball track cannot observe how high the ball went, so never read height off these curves.
Click any arc to jump the video to that shot. Drag to rotate, scroll to zoom.

Interactive 3D court visualization — drag to rotate · scroll to zoom · click a preset to snap to a view angle.

📊 Match Timeline

⚡ Ball Speed Analytics ⓘHow speed is computed
Ball positions are tracked frame-by-frame via TrackNet-style temporal differencing. Per-step displacement is converted to real-world mph using the auto-detected court corners as a known 44 ft (13.4 m) reference scale. Peak speed uses the 95th percentile and is capped at a realistic pickleball ceiling (~62 mph) to reject ball-detection glitches. Without court calibration treat speeds as approximate.

🏠 Court Zone Dwell Time ⓘThe #1 pickleball metric
Time each player spent in three court zones: kitchen (within 7 ft of net), transition (mid-court), and baseline (back). The kitchen is where dinks are won and points are decided — high kitchen % usually correlates with control. Computed from feet positions sampled at 4 Hz throughout the match.

🏓 Rally Analytics ≈ ESTIMATE ⓘHow rallies are grouped
Shot events (paddle contacts) detected by wrist-velocity peaks are grouped into rallies — a gap of more than 2.5 seconds between contacts ends a rally. Rallies with fewer than 2 contacts are discarded as serves or noise. These are estimates: the counts depend on shot detection + player tracking, which are heuristic today, so totals can drift and per-rally player tags are the least reliable.

Grouped from detected shot events — approximate. Counts drift if shots are over/under‑detected, and the per‑rally player tags are the least reliable part (player‑tracking limitation).

🤚 Swing Distribution ⓘHow swings are classified
At each shot contact moment, the player's 17-keypoint pose is analyzed: wrist position relative to body centerline (FH vs BH), height vs nose (overhead), and height vs hip (dink/low). This uses real biomechanics rather than spatial heuristics — more accurate and camera-orientation agnostic.

🎯 Error Analysis ⓘHow errors are detected
Every detected ball bounce is tested against the court polygon (auto-detected corners). Bounces strictly outside the boundary, after a small margin for detection slack, are scored as unforced errors. Each error is classified by zone (long, wide, sideline) so coaches can see WHERE errors happen — not just how many.

💡 AI Coach Insight ⓘHow it's generated
A large language model (GPT-class) reads each player's metrics (court positioning, swing mix, reaction times, biomechanics) and writes one independent coaching paragraph per player. When the LLM is unavailable, a rule-based fallback produces the same per-player breakdown. Regenerates each time the video is reanalyzed.

📊 Statistics ⓘWhat's included
Raw aggregate metrics computed over all analyzed frames: total rally count, average & max rally duration, shot counts per player, average court positioning zone, and paddle-speed histogram. These are the source values used by all other modules — export via PDF for offline review.