Shot detection and outcome recognition on complete training videos
We evaluated PerformStars on complete, unedited pool training videos and compared the automatic output with a detailed manual review.
The published measurements come from pool training sessions and evaluate the core computer vision algorithms used by PerformStars.
Latest test · September 14, 2026
124 shots in a new validation session
Shot Detection Recall
The share of 124 real shots detected automatically.
Outcome Accuracy
Accuracy when the system confidently determined the shot outcome.
Decision Coverage
The share of detected shots for which the system determined an outcome.
This version improves trajectory recovery when a fast-moving ball briefly disappears from view and matches cue-ball motion with object-ball contact, including bank shots. To estimate speed and spin, the system compares possible motion scenarios under the actual table conditions.
The sessions use different source videos, so these are separate validation measurements rather than a direct version comparison. Detection precision and F1 were not calculated separately for the September 14 test.
Watch source video ↗Shot Detection Recall
The system automatically detected 65 of 72 real shots.
Shot Detection Precision
65 of 68 events detected by the system were real shots.
F1 Score
Combined balance between detection precision and recall.
Outcome Accuracy
57 of 58 automatic shot outcome decisions were correct.
September 3, 2026
Automatic Shot Detection
The complete session contained 72 real shots according to the final manual review. PerformStars automatically generated 68 shot candidates. Of these, 65 were confirmed as real shots and 3 were false detections.
The seven shots found during manual review are not included in the automatic detection result.
September 3, 2026
Made / Miss Recognition
We separately evaluated whether PerformStars could correctly determine the result of shots it had already detected.
Accuracy when a decision was made
57 of 58 automatic made/miss decisions were correct.
Decision Coverage
The system produced a made/miss decision for 58 of 65 automatically detected real shots.
PerformStars uses a conservative decision strategy. When the available signals are not strong enough, the system can return an unknown result instead of forcing a made/miss classification.
Methodology
In each test, PerformStars processed a complete training video using its normal automatic analysis pipeline. The session was then reviewed manually to establish ground truth for every real shot and its final result.
Automatic detections and the original automatic made/miss decisions were compared with this manual reference. Shots recovered only during manual review were not counted as automatic detections, and no manual result was substituted for an automatic result.
Source video
Complete unedited training session.
Automatic analysis
PerformStars processed the video without manual intervention.
Ground-truth review
All real shots and outcomes were manually verified.
Comparison
Original automatic results were compared with the verified session.
Verify the tests yourself
Review the source videos for both tests. The processed PerformStars workout is also available for the September 3 session.
Latest source video
Complete unedited training video. The processed workout has not yet been published.
Watch videoExplore the processed workout
Open the actual PerformStars workout and review the detected shots and their results.
Open workout dataValidation Summary
September 14, 2026 · 124 shots
September 3, 2026 · 72 shots
These results describe specific recorded pool training sessions and should not be interpreted as a universal accuracy guarantee for every video, table, camera angle, lighting condition, recording quality, or style of play.
Test PerformStars on your own session
Upload a practice video and see how PerformStars turns real footage into structured shot data and coaching analysis.