Standard 6 stops to get here

Keypoint Detection

Finding specific, repeatable points in an image: distinctive spots for matching two views, or named landmarks such as a person's elbows and knees.

Your route here

6 stops · basics first

Machine Learning ✓ understood
Building systems that learn patterns from data instead of following hand-written rules, getting better at a task as they see more examples.
Neural Network ✓ understood
A model built from layers of simple units, each taking a weighted sum of its inputs and applying a nonlinearity; training adjusts the weights so the whole stack maps inputs to useful outputs.
Deep Learning ✓ understood
A subset of machine learning that uses neural networks with multiple layers (deep neural networks) to learn hierarchical representations of data.
Computer Vision ✓ understood
The field of AI that gets computers to extract meaning from images and video: what is in them, where it is, and how it moves.
Convolution ✓ understood
A mathematical operation that applies filters/kernels to input data to extract features like edges, textures, and patterns.
Convolutional Neural Network ✓ understood
A neural network that scans images with small learned filters, reusing the same weights at every position to build up from edges to whole objects.
Keypoint Detection · you are here ✓ understood

Keypoint detection outputs a set of points rather than a label or a box. The term covers two related jobs in computer vision, and it helps to know which one someone means.

Interest points for matching

The classic sense is finding spots that can be found again in another photo of the same scene: corners, blobs, distinctive texture. Each detected point gets a descriptor, a short vector summarizing its neighbourhood, so points can be matched across images. Lowe’s SIFT is the landmark example. Its features are invariant to image scale and rotation and stay matchable under changes in viewpoint, noise and lighting. Matched keypoints drive panorama stitching, visual tracking and 3D reconstruction.

Semantic landmarks

The deep-learning sense is locating named parts: the joints of a human skeleton, the corners of the eyes and mouth, the wheels of a car. This is the engine of pose estimation. A CNN typically predicts one heatmap, or confidence map, per keypoint type, and the peak of each map is the predicted location.

With several people in a frame there are two strategies. Top-down methods run an object detector first and then find keypoints inside each person box, so the cost grows with the number of people. Bottom-up methods detect every keypoint in the image at once and then group them into individuals. OpenPose took the bottom-up route: it predicts confidence maps for body parts plus “part affinity fields” that link parts into limbs, and it runs in real time however many people are present.

Where it sits

Before this

Computer Vision Convolutional Neural Network

Keypoint Detection

Leads to

Nothing yet: a destination in its own right.

Explore nearby

Vision & Multimodal Pose Estimation Detecting body keypoints to determine human pose from images or video. Vision & Multimodal Object Detection Finding every object of interest in an image and giving each a class label, a confidence score and a bounding box. Neural Networks Feature Map The output of applying a convolutional filter to an input, representing detected features at various spatial locations. Vision & Multimodal 3D Reconstruction Creating 3D models from 2D images using geometry and deep learning. Vision & Multimodal Optical Flow Estimating motion patterns between video frames by tracking pixel movements. Vision & Multimodal Facial Recognition Identifying or verifying people from face images using deep learning.

Sources

David G. Lowe, "Distinctive Image Features from Scale-Invariant Keypoints" . International Journal of Computer Vision, 2004 (SIFT)
Cao, Simon, Wei and Sheikh, "Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields" . CVPR 2017 (OpenPose)