Weakly-Supervised Multi-View Image-Based Face Reconstruction: a Novel Method
Yang Han, Chunya Zhao, Lele Guan · 2023
Current research mostly focuses on single-image-based facial modeling reconstruction, which may not accurately capture facial depth information. Therefore, this paper proposes using multi-views to reconstruct facial models of the target. Firstly, we devise a weakly supervised feature extraction network to fuse features from multiple views and generate consistent facial models, an approach less explored in current research. Secondly, we introduce an attention mechanism in the network to better extract essential facial features, adapting to various view. Additionally, to address the issue of unobservable facial landmarks in side-view images, we propose an adaptive landmark weighting method. Instead of using invisible landmarks, we employ visible landmarks for model generation. Experimental results show that our method achieves approximately 5.56% improvement in RMSE compared to other methods. With these contributions, this paper presents a novel solution for multi-view facial modeling.