Fine-grained Human Motion Understanding with Language Models
FiGMo is an LLM-based model for fine-grained motion understanding that treats human motion as a sequence of skeletal poses with explicit timestamps. It overcomes previous limitations by grounding every pose in time, allowing the model to reason accurately about the order, duration, and rhythm of actions. The model uses a unified encoder for 2D and 3D poses and is trained with diverse supervision to achieve state-of-the-art results.