JD AI Research Institute Wins CVPR 2018 LIP Human Pose Estimation Competition

On June 11, 2018, the JD AI Research Institute's Computer Vision and Multimedia Laboratory secured the champion titles in both the single‑person and multi‑person human pose estimation tracks of the CVPR 2018 Look Into Person (LIP) competition, and also earned the runner‑up position in the single‑person pose parsing track.

The LIP competition, now in its second year, attracted top research groups from universities such as Carnegie Mellon, UCLA, UC Berkeley, Johns Hopkins, and NUS, as well as institutions like the University of Illinois, Ohio State, Tsinghua, Nanjing, and many others. Winners were invited to present their solutions at the IEEE CVPR 2018 conference in Salt Lake City.

For the single‑person pose estimation track, the JDAI‑Human team improved existing multi‑scale fully‑convolutional networks (e.g., Cascading Pyramid Network, Stacked Hourglass, RMPE) by enhancing kernel detail perception, applying extensive image augmentation, expanding data, adjusting hard‑example mining key‑point classes, and employing focal loss. The fused model achieved 90.9% accuracy, setting a new record.

In the newly added multi‑person pose estimation track, the team refined a top‑down framework that combines person detection with single‑person key‑point estimation. Improvements included a stacked hourglass model with pyramid residual modules, a parametric Pose NMS to reduce redundant detections, and a strategy for selecting hard key‑points and loss weights. This approach reached 72.2% accuracy, 2.3% ahead of the runner‑up.

The laboratory’s technologies in human behavior analysis, face recognition, and person re‑identification will soon be available on JD AI’s NeuHub platform (http://neuhub.jd.com/), supporting applications such as borderless retail, unmanned stores, warehouses, and AR/VR marketing. JD AI continues to collaborate with partners to bring these advanced solutions to real‑world business scenarios.