Darts and AI: What Artificial Intelligence Really Detects in Your Throw mydart Darts and AI: What Artificial Intelligence Really Detects in Your Throw ========================================================================= Behind AI dart detection sits one concrete technique: pose estimation. What gets measured, where the limits are and how to spot reliable detection. Pose-Tracking Echtzeit 8+ Metriken [ mydart ](https://mydart.app/en) [Features](https://mydart.app/en/features) [Training](https://mydart.app/en/training) [Price](https://mydart.app/en/pricing) [Blog](https://mydart.app/en/blog) EN [ DE Deutsch ](https://mydart.app/de/blog/dart-ki) EN English [ Login ](https://mydart.app/login) [ Features ](https://mydart.app/en/features) [ Training ](https://mydart.app/en/training) [ Price ](https://mydart.app/en/pricing) [ Blog ](https://mydart.app/en/blog) [ Login ](https://mydart.app/login) 1. [Home](https://mydart.app/en) 2. [Blog](https://mydart.app/en/blog) 3. Darts and AI: What Artificial Intelligence Really Detects in Your Throw Analysis 9 min read • 19. August 2026 Darts and AI: What Artificial Intelligence Really Detects in Your Throw ========================================================================= Behind AI dart detection sits one concrete technique: pose estimation. What gets measured, where the limits are and how to spot reliable detection. What AI actually means in darts ------------------------------- The term AI gets used inflationarily in sport and covers very different things. In darts it comes down essentially to a single technique: pose estimation, the detection of human body posture in images. Everything that happens afterwards is ordinary mathematics. That is not a criticism, quite the opposite. Pose estimation solves a problem that was practically unsolvable with classical programming. Nobody can write down rules for how to recognize an elbow joint in a photo. A neural network that has seen hundreds of thousands of annotated images can. So AI dart detection is at its core the job of deriving the position of your joints from camera frames. How pose estimation works ------------------------- The model receives an image and returns coordinates. In the common approaches those are 33 points: nose, eyes, shoulders, elbows, wrists, fingers, hips, knees, ankles and further references. For each point the model also returns a confidence value that states how certain it is about that position. Such a network was trained on enormous image collections in which humans marked joint positions by hand. From those examples it learns which visual patterns typically indicate a wrist and which do not. What is remarkable is that it never explicitly learned anatomy. It internalized statistical relationships between pixels and annotations. When this runs on video instead of a single frame, each point produces a motion trail over time. At that moment the AI's work ends and the actual evaluation begins. From points to statements ------------------------- An elbow angle is nothing more than the angle between two segments: shoulder to elbow and elbow to wrist. Once the three coordinates exist, the calculation is pure geometry and produces exact results. The same holds for speeds, which follow from the change in position between two frames, and for spread, which is calculated across many throws. This separation is important to understand, because it explains where errors can arise and where they cannot. The geometry always computes correctly. The only uncertain step is the first one, the question of whether the AI really located the wrist at the wrist. When an evaluation looks implausible, the cause is almost always the detection and almost never the calculation. What artificial intelligence cannot do in darts ----------------------------------------------- Three things are regularly expected of dart AI that it cannot deliver. ### It does not know the ideal technique A model that detects joint positions has no concept whatsoever of what a good throw looks like. It has seen people in images, not judgements. Every statement about whether an angle is favourable comes from rules humans defined afterwards, or from comparison with your own earlier throws. Anyone claiming their AI knows the perfect throwing technique is selling an opinion as a measurement. ### It does not understand your intent Whether you deliberately aimed at treble 20 or at double 16 does not follow from the movement. AI can identify deviations from your own pattern, but it cannot judge whether a deviation was a mistake or a conscious adjustment. ### It does not replace an explanation The measurement tells you that your elbow angle dropped by 6 degrees in the second half of the session. Why that happened, whether from fatigue, concentration or a change in stance, is in no data set. That step is still made by a human, either you or a coach. The article [does mydart replace a real coach](https://mydart.app/en/blog/does-mydart-replace-a-real-coach) deals with exactly this boundary. On the device or in the cloud ----------------------------- One decision with noticeable consequences concerns where the computation happens. If detection runs on a server, the video has to be transferred there. That means waiting time, dependence on the connection and above all: your footage leaves your device. If it runs locally on a phone or tablet, neither applies. Modern devices handle pose estimation in real time because the models are now small enough. The price is slightly lower accuracy than the largest server models offer, the gain is immediate feedback and the fact that no recording of your throwing room is uploaded anywhere. **mydart** works locally. Camera frames are processed on the device, and only the calculated metric values are transmitted. For most players that is the more important factor, because a throwing video always contains footage of your own home as well. How to spot reliable detection ------------------------------ Not every app advertising AI produces usable values. These points are a good test grid: - **A visible skeleton.** If the app overlays the detected points on the camera image, you can see for yourself whether they sit on your joints. Systems that hide this view make checking impossible. - **Handling of uncertainty.** A good evaluation discards throws where detection was unreliable instead of folding questionable values into the average. - **Traceable metrics.** If an app reports a "throw score" of 78 without explaining what it consists of, you cannot train with it. - **Reaction to camera changes.** Deliberately set the camera up differently and see whether the values shift heavily. A system that handles perspective correctly should stay robust. Common sources of error in practice ----------------------------------- Most detection problems have mundane causes. Loose dark clothing against a dark background makes it hard for the model to find the contour of the arm. Backlight from a window behind you turns your body into a silhouette. With the camera placed head on, upper and lower arm overlap in the image and the elbow angle becomes nearly impossible to determine. All three are solved without any technical effort: camera to the side at shoulder height, even light from the front or the sides, and clothing that stands out from the background. That removes by far the largest share of detection errors before they occur. What is likely to change in the coming years -------------------------------------------- The models will get smaller and more accurate, that is the most stable prediction in this field. In practice that means higher frame rates and better detection of the hand, currently the hardest area because fingers are small and occlude each other. For analyzing the grip and the moment of release that would be a clear step forward. What will not change is the basic division of labour: AI supplies measurements, while interpretation and the decision about what to change in training stay with you. Which values are meaningful in the first place is covered in the article on [throw analysis and its metrics](https://mydart.app/en/blog/throw-analysis-metrics). Related reading --------------- [ Analysis Detecting a Dart from a Photo: How Camera Recognition Works in Darts Detecting a dart from an image can mean two different things. This article explains how image recognition works, where its limits are and why motion reveals more than a still frame. Read more ](https://mydart.app/en/blog/detect-a-dart-from-a-photo) [ Analysis Throw Analysis in Darts: The Metrics That Actually Matter Throw analysis produces numbers, but which of them help in training? Elbow angle, release timing, follow-through, rhythm and consistency explained, including their limits. Read more ](https://mydart.app/en/blog/throw-analysis-metrics) [ Analysis Dart Motion App: Understanding Your Throw in Slow Motion Dart motion analysis makes visible what happens in 400 milliseconds. What to watch in slow motion, why frame rate decides accuracy and when an app beats a phone recording. Read more ](https://mydart.app/en/blog/dart-motion-app) Follow us [ ](https://www.facebook.com/profile.php?id=61588643532421) [ ](https://www.instagram.com/mydart.app/) [ ](https://www.tiktok.com/@mydart.app) [ mydart ](https://mydart.app/en) Improve your throwing technique with real-time feedback and pose tracking. Now available as an app [ ![Download mydart on the App Store](https://mydart.app/storage/apple-black-en.svg)](https://apps.apple.com/de/app/mydart-app/id6762560486?ct=organic-footer-mydart-web&mt=8) [ ![Get mydart on Google Play](https://mydart.app/storage/android-black-en.png)](https://play.google.com/store/apps/details?id=com.mydart.android&referrer=utm_source%3Dmydart-web%26utm_medium%3Dreferral%26utm_campaign%3Dorganic%26utm_content%3Dfooter) Product - [Features](https://mydart.app/en/features) - [Training](https://mydart.app/en/training) - [Price](https://mydart.app/en/pricing) - [Blog](https://mydart.app/en/blog) - [Partnerships](https://mydart.app/en/partnerships) Account - [Login](https://mydart.app/login) Legal - [Imprint](https://mydart.app/en/imprint) - [Privacy](https://mydart.app/en/privacy) - [Terms](https://mydart.app/en/terms) © 2026 mydart. All rights reserved.