DrawMotion: Generating 3D Human Motions by Freehand Drawing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Text-to-motion generation models struggle with precise user intent due to limitations of textual descriptions.
• Existing methods lack intuitive control for fine-grained motion specification.

Purpose of the Study:

• Introduce DrawMotion, an efficient diffusion-based framework for multi-condition text-to-motion generation.
• Enable users to control motion generation through both semantic (text) and spatial (hand-drawing) conditions.
• Improve the accuracy and user efficiency in generating human motions.

Main Methods:

• Developed a novel hand-drawing condition for intuitive spatial control, complemented by a 2D trajectory condition.
• Proposed a Multi-Condition Module (MCM) for efficient fusion of text and drawing inputs within the diffusion process.
• Implemented training-free guidance using MCM's continuous feature space to align generated motions with user intent while preserving fidelity.

Main Results:

• Quantitative experiments and user studies validated the effectiveness of the hand-drawing approach.
• The freehand drawing method reduced user time by approximately 46.7% for motion generation aligned with user imagination.
• DrawMotion demonstrated improved control and fidelity in generating human motions based on combined conditions.

Conclusions:

• DrawMotion offers an efficient and intuitive solution for text-to-motion generation by incorporating a hand-drawing condition.
• The Multi-Condition Module effectively fuses diverse inputs, enhancing control and reducing computational complexity.
• This framework significantly improves user experience and efficiency in creating desired human motion sequences.