Fine-grained Motion Retrieval via Joint-Angle Motion Images and Token-Patch Late Interaction

This paper introduces a novel approach to text-motion retrieval by addressing the limitations of traditional global-embedding paradigms. Most existing methods compress 3D human motion sequences and textual descriptions into single global vectors, which is computationally efficient but inevitably discards fine-grained local correspondences. To overcome this information bottleneck, the authors propose replacing global embeddings with a structurally grounded, fine-grained late interaction mechanism.

The core innovation relies on explicitly decoupling local joint movements from the body's global trajectory using a joint-angle-based motion representation. These joint-level features are mapped into a structured 224x224 pseudo-image, making them compatible with pre-trained Vision Transformers. For the retrieval process, the framework utilizes MaxSim, a token-wise late interaction operator that matches text tokens to motion patches, which is further stabilized by a Masked Language Modeling regularization on the text encoder.

Extensive experiments conducted on the HumanML3D and KIT-ML datasets demonstrate that this approach consistently outperforms state-of-the-art text-motion retrieval baselines. Furthermore, the token-to-patch matching produces highly interpretable correspondence maps that transparently align textual semantics with specific body joints and temporal phases. This level of transparency and granularity provides a robust foundation for downstream applications like language-driven motion generation and localized editing.