HumanSkill-4D queries animatable Gaussian avatars without training
Original title: HumanSkill-4D: Training-Free Language Querying of Animatable Gaussian Avatars in Space and Time
The training-free method beat baselines on part segmentation and 3D and 4D referring segmentation using three new human-annotated benchmarks.
HumanSkill-4D is a training-free framework for avatar referring segmentation: selecting regions of animatable Gaussian avatars from language queries, including pose or motion intervals. A shared multimodal large language model and geometric operators use a semantic memory of body-part and garment categories and a motion memory of driving joint trajectories. Category choices and operator programs are meant to make selections traceable. On three human-annotated benchmarks it outperformed baselines in part segmentation and 3D and 4D referring segmentation, with compact memories and minutes of setup per avatar.
- Queries can name a pose or motion and select matching time intervals.
- The method targets region ambiguity and similar-looking fine-grained motions.
- Setup is described as only minutes per avatar.
- Operators can be composed or added for new motion conditions.
From the original article
1 min read at huggingface.co
We introduce avatar referring segmentation, which selects regions of animatable Gaussian avatars from language queries. A query can also specify a pose or motion to select matching intervals. This lets users interact with avatars for local editing and motion inspection.
By unsu Kim, Naeun Kim, Dongyoon Han · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.