MorekerLess – Better pose estimation with less work

Markerless motion capture is becoming the new norm for kinematic analysis, but some challenges remain. One of them lies in 2D pose estimation: anatomical keypoints could be labeled more accurately and should be more numerous for certain parts of the body, such as the spine. However, manually creating a new dataset or modifying annotations is extremely costly, as hundreds of thousands of images are generally required. We introduce MorekerLess, a pipeline for automatically labeling and training on any keypoint set. Results show that additional keypoints are correctly detected, and joint center estimation is improved by 0.2-0.4 cm over the current state-of-the-art. Our pipeline and models will be released at: http://github.com/davidpagnon/morekerless.