Abstract
We present a fast bottom-up method that jointly detects over 100 keypoints onhumans or objects, also referred to as human/object pose estimation. We modelall keypoints belonging to a human or an object -- the pose -- as a graph andleverage insights from community detection to quantify the independence ofkeypoints. We use a graph centrality measure to assign training weights todifferent parts of a pose. Our proposed measure quantifies how tightly akeypoint is connected to its neighborhood. Our experiments show that our methodoutperforms all previous methods for human pose estimation with fine-grainedkeypoint annotations on the face, the hands and the feet with a total of 133keypoints. We also show that our method generalizes to car poses.