OpenPose explain

It breaks down the task into Body Keypoint Detection, Hand Keypoint Detection, Facial Keypoint Detection, and Multi-Person Support.

Core idea.

The core component of OpenPose is the body keypoint detection module, which predicts the 2D locations of key body joints like elbows, knees, shoulders, and hips. This is achieved through the following steps:

1. A convolutional neural network (CNN) extracts features from the input image.
2. The extracted features are fed into two parallel CNN branches:

• One branch predicts 18 confidence maps, each representing the confidence of a specific body part at each pixel location.
• The other branch predicts a set of 38 Part Affinity Fields (PAFs) that encode the degree of association between different body parts.

3. The confidence maps and PAFs are parsed to construct bipartite graphs, where nodes represent detected body parts and edges represent their associations.
At this step, NMS plays a crucial role in OpenPose by filtering out duplicate or low-scoring body part detections from the confidence maps.

Get noplaxochia’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

4. These bipartite graphs are processed to assign body parts to individual people, generating complete human pose skeletons.