Kenkun Liu (刘垦坤)

An LLM-based framework that unifies text-to-motion generation and motion-to-text understanding with motion-consistent chain-of-thought data and GRPO post-training.

A DETR-style model that represents 3D Gaussians as unified world-space queries and updates them via multi-view deformable cross-attention for sparse-view reconstruction.

An end-to-end, one-stage system that uses temporally consistent queries to capture every person's whole-body motion from monocular video without separate detectors or trackers.

A deformable-Transformer method that directly models each query as a 3D Gaussian, removing pixel and point-cloud correspondence for compact, efficient single-view reconstruction.

A large-scale multi-view human dataset with 4,500 identities, 9,000 outfits, 60,000 motion sequences, and 645 million richly annotated frames.

A unified benchmark for fairly evaluating human NeRFs across challenging scenes, with cross-subject tests and analyses of animatability and generalizability.

A 3D pose-estimation GCN that adds a global node and heterogeneous multi-task learning to combine joint-local features with global pose attributes.

A systematic study that derives and evaluates eight GCN weight-sharing schemes for modeling distinct body-joint relationships in 3D pose estimation.