Keypoint Detection

Keypoint detection teaches a vision system to locate meaningful points in an image: the corner of an eye, a fingertip, the center of a knee, or the tip of a tool. Rather than merely deciding that a person or object is present, the system identifies where its important parts are.

What the model predicts
For each expected landmark, a model produces an image coordinate such as (x, y), usually along with a confidence score. A pose-estimation model, for example, can predict 17 body keypoints including shoulders, elbows, hips, knees, and ankles. Modern neural networks commonly create one heatmap per keypoint: a bright peak marks the location the network considers most likely. Some models instead directly regress numerical coordinates.

How it is trained and used
Training requires images labeled with landmark positions. The model compares its predicted locations with those annotations and learns to reduce the distance between them. A practical keypoint pipeline usually includes:

• finding a person, face, hand, or object region;
• predicting its landmarks within that region;
• discarding low-confidence or occluded points; and
• connecting points into a skeleton, shape, or geometric measurement.

For instance, MediaPipe Pose estimates body landmarks from video frames, enabling fitness apps to count repetitions or assess joint angles. Facial keypoints support face alignment before recognition; eye and mouth locations let the recognizer compare faces in a consistent orientation.

Why precise points matter
Keypoints unlock decisions that bounding boxes cannot make. An autonomous vehicle can distinguish a pedestrian standing still from one stepping into the road by tracking body joints across frames. In medical images, landmarks can measure organ dimensions or guide alignment between scans. On a production line, keypoints on a connector or screw can reveal whether a part is rotated, bent, or assembled incorrectly. This term is distinct from classical “interest-point” detectors such as OpenCV’s goodFeaturesToTrack, which find visually distinctive corners without knowing that they represent an eye, elbow, or other named body part. Semantic keypoint detection combines that geometric precision with an understanding of what each point represents.