CustomDancer: Customized Dance Recommendation by Text-Dance Retrieval

Dance serves as both a cultural cornerstone and a medium for personal expression, yet the exponential growth of online dance content has intensified the need for personalized retrieval systems. While text-based dance retrieval offers a promising avenue for users to articulate preferences through natural language, existing methods face critical challenges: aligning abstract textual semantics with motion kinematics, reconciling cross-modal heterogeneity, and overcoming the scarcity of annotated data. To address these gaps, we presentCustomDancer, a novel framework that bridges music, motion, and text through multimodal fusion. Our work introducesTD-Data, the largest open-source dataset for text-dance retrieval, comprising about 4,000 professionally annotated clips across 22 genres, refined via a three-stage pipeline integrating dance preprocessing, expert annotation, and AI-driven text refinement. The CustomDancer framework employs domain-adapted encoders for text, music, and motion, alongside a hybrid fusion module combining additive and multiplicative interactions to align auditory and kinematic features. Experiments demonstrate the superiority of our framework and effectiveness of our datasets. This work establishes foundational benchmarks for personalized dance interaction while addressing key challenges in cross-modal retrieval.