Detect a Dart from a Photo: What Image Recognition Can Do

Two very different questions behind one search

People who want to detect a dart from an image usually mean one of two things. The first group has a dart in hand, or a photo of one, and wants to know which model it is. Which manufacturer, which weight, which barrel shape. The second group wants something entirely different: they want a camera to recognize where the dart landed in the board, or how the hand threw it.

Technically these two tasks have little in common, even though both get searched with phrases like "identify a dart from a photo" or "dart detection". One is a classification problem backed by a product database, the other is a question of object localization and motion capture. This article covers both, because confusing them leads to the wrong expectations of any app.

How a computer recognizes anything in an image

To software, a photo is initially nothing but a table of numbers. Every pixel carries three values for red, green and blue. That is a long way from "there is a dart". The bridge between numbers and meaning is built by neural networks trained on very large numbers of example images.

During training the network sees thousands of images in which a human has marked where the target object sits. It does not learn rules in any human sense, it learns patterns: certain edge shapes, brightness transitions, typical outlines. After enough examples it can point at a new, never seen image and state where a dart probably is and how confident it is.

There are three levels of precision here. Classification only states whether something appears in the image. Object detection adds a rectangle around it. Segmentation marks the outline pixel by pixel. To answer where a dart sits in the board you need object detection at minimum, ideally segmentation, because a few millimetres at the tip decide between one segment and its neighbour.

Detecting a dart from a photo: why a single frame is hard

A dart is an awkward object for image recognition. It is thin, often shiny metal, and it sits at an angle in a surface that is itself made of lines and coloured segments. Three problems come up again and again.

Occlusion

When three darts sit close together, the front one hides the tip of the one behind it. A human infers the tip position from the visible shaft. A model that only looks at single frames does not have that information and guesses.

Perspective

A photo taken from below distorts the board into an ellipse. Without correcting that distortion, every position estimate is off. Systems that work reliably first detect the board geometry, transform it into a top down view, and only then determine the landing position.

Light

Metal barrels reflect. Depending on the lamp, a bright streak appears exactly where the model expects an edge. Poor or changing light is the most common cause of detection errors in practice, more common than any weakness in the model itself.

From still image to motion

As soon as a video replaces a single photo, the situation changes considerably. At 30 or 60 frames per second a motion trail emerges, and that trail carries information no still image can provide. A flight path can be followed across several frames, and brief occlusions are bridged by the frames before and after.

The second big advantage is not about the dart, it is about the person. When it comes to why a throw misses, the landing position is only the result. The cause lies in the movement: in the elbow angle, in the timing of the release, in how the arm swings out. This is exactly where pose tracking works. Instead of looking for the dart, the software recognizes the body and places 33 reference points on shoulder, elbow, wrist, hip and legs.

Angles, speeds and deviations emerge from how those points move over time. That is the approach mydart takes: not photographing the board, but filming the thrower and translating the motion into numbers. How those numbers are produced and what they mean is covered in detail in the article on dart throw analysis.

Identifying a dart: which model is that in the board?

The other meaning of dart detection aims at product identification. You see a barrel you like on someone else's throw, take a photo and want to know which model it is. Technically this is a similarity search: the system compares your photo against a database of known barrels and returns the closest matches.

The catch is the database. There are thousands of barrel models, many differ only in the knurling or by a couple of grams, and weight is fundamentally invisible in a photo. Image recognition can narrow down the shape and the groove pattern, but it cannot resolve the difference between 21 and 23 grams of the same series. Realistically you get a shortlist, not a definitive answer.

In practice that means this: if you want to know which dart someone throws, asking is faster and more reliable than any app. And if you want to know why they hit better than you do, the model name helps you very little anyway. The barrel accounts for a small part of the result, the movement for by far the larger one.

What makes a good recording

Whichever system you use, the recording decides the quality of the detection. These points make the biggest difference in practice:

• Even light. Two light sources from the sides beat one bright lamp from above. Hard shadows and reflections are the most common reason for detection errors.
• A fixed camera position. A tripod or mount beats any handheld recording. If the frame shakes, the reference moves and every measured value moves with it.
• A calm background. The less motion and structure behind you, the cleaner the software separates person from surroundings.
• Side position for motion analysis. If your throwing motion is the subject, the camera belongs beside your throwing arm, roughly at shoulder height. From the front, elbow angle and pull back are nearly impossible to measure.
• The whole body in frame. Even if only the arm interests you: stance and weight shift are part of the movement and should be visible.

What mydart detects and what it does not

mydart uses the camera of your phone or tablet to capture your throwing motion and translate it into measurable values. What it detects are body points and their movement over time, and from those come elbow angle, follow-through, rhythm and consistency across a session. The evaluation runs locally on the device, no video frames are sent to a server.

What it does not detect is the model of your barrel, and the app does not score the board either. That is a deliberate decision: the landing position tells you that a throw missed, the motion data tells you why. For the question of why, the body is the more interesting image source than the board.

If you want to know which metrics come out of this and how to read them, the overview of the dart throw analysis app covers it.