"UniHM"

"Human motion synthesis in complex scenes requires rich, multimodal data integrating human motion, static environments, movable objects, and natural language descriptions. Existing datasets are often limited in scale, modality diversity, or scene complexity, constraining the development of scene-aware motion models. To address this, we present the UniHM dataset, a large-scale, unified dataset for human motion and human-object interaction (HOI) in 3D scenes. UniHM combines sequences from HumanML3D, OMOMO, and the newly introduced Lingo Human-Scene-Interaction dataset, standardized to a global SMPL representation and enriched with scene voxels, object point clouds, and multi-text annotations. The dataset contains 44,962 sequences of up to 300 frames each, supporting both Text-to-Motion and Text-to-HOI tasks. By providing comprehensive motion, scene, and object modalities with consistent annotations, UniHM enables the development and evaluation of robust, scene-aware motion synthesis models."