DepthDance: Complex-pose Human Image Animation with Appearance-agnostic Depth Guidance

Authors: Xi, Y., Xu, Z., Wang, Z., Yang, X., Lan, J., Zhang, J.J., Chen, M.

Journal: Proceedings 2025 IEEE Cvf International Conference on Computer Vision Workshops Iccv W 2025

Publication Date: 01/01/2025

Pages: 1947-1957

DOI: 10.1109/ICCVW69036.2025.00204

Abstract:

The field of human image animation has witnessed significant progress. However, modeling human motion with complex poses remains a formidable challenge. The existing methods predominantly mainly rely on skeletons representations for pose control, yet pose estimators often fail to provide reliable guidance under complex scenarios. Given that depth maps exhibit greater robustness in challenge scenarios, we propose DepthDance to explore how to utilize depth maps as the control signal in this work. However, a novel challenge emerges from the fact that depth maps introduce fine-grained appearance information, which complicates the transfer of motion across different identities. We propose the Auxiliary Depth Information Injection Method (ADI² M), designed to mitigate appearance leakage inherent in depth maps. To enhance the model's generalizability across a range of motion complexities and to leverage pre-trained skeleton-based models, we introduce Pose Curriculum Learning Strategy (PCurLS) for more fine-grained control. Extensive experiments validate the effectiveness of our approach, especially in scenarios involving challenging human poses where traditional control signals often fail.

Source: Scopus