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

TL;DR

This paper introduces a novel interpretable approach for text-to-motion retrieval that leverages joint-angle motion images and token-patch late interaction, significantly improving accuracy and interpretability over existing global-embedding methods.

Contribution

It proposes a joint-angle-based motion representation and a token-wise late interaction mechanism with regularization, enabling fine-grained, interpretable text-motion alignment.

Findings

Outperforms state-of-the-art retrieval methods on HumanML3D and KIT-ML datasets.

Provides interpretable, fine-grained correspondences between text and motion.

Enhances robustness with Masked Language Modeling regularization.

Abstract

Text-motion retrieval aims to learn a semantically aligned latent space between natural language descriptions and 3D human motion skeleton sequences, enabling bidirectional search across the two modalities. Most existing methods use a dual-encoder framework that compresses motion and text into global embeddings, discarding fine-grained local correspondences, and thus reducing accuracy. Additionally, these global-embedding methods offer limited interpretability of the retrieval results. To overcome these limitations, we propose an interpretable, joint-angle-based motion representation that maps joint-level local features into a structured pseudo-image, compatible with pre-trained Vision Transformers. For text-to-motion retrieval, we employ MaxSim, a token-wise late interaction mechanism, and enhance it with Masked Language Modeling regularization to foster robust, interpretable text-motion alignment. Extensive experiments on HumanML3D and KIT-ML show that our method outperforms state-of-the-art text-motion retrieval approaches while offering interpretable fine-grained correspondences between text and motion. The code is available in the supplementary material.