Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CSANet: Channel Spatial Attention Network for Robust 3D Face Alignment and Reconstruction

Yilin Liu, Xuezhou Guo, Xinqi Wang, Fangzhou Du

TL;DR

CSANet addresses robust 3D face alignment and reconstruction from single 2D images, focusing on occlusion and lighting challenges. It augments a lightweight bottleneck backbone with Spatial Group-wise Enhancement and Coordinate Attention to refine features, and uses a joint Wing Loss and WPDC objective to stabilize 3DMM parameter learning. The approach yields superior accuracy, especially under large poses, and demonstrates faster training compared to baselines like 3DDFA, with competitive reconstruction quality on AFLW/AFLW2000-3D. This work provides a practical, attention-based framework for efficient, robust 3D face modeling suitable for real-world applications.

Abstract

Our project proposes an end-to-end 3D face alignment and reconstruction network. The backbone of our model is built by Bottle-Neck structure via Depth-wise Separable Convolution. We integrate Coordinate Attention mechanism and Spatial Group-wise Enhancement to extract more representative features. For more stable training process and better convergence, we jointly use Wing loss and the Weighted Parameter Distance Cost to learn parameters for 3D Morphable model and 3D vertices. Our proposed model outperforms all baseline models both quantitatively and qualitatively.

CSANet: Channel Spatial Attention Network for Robust 3D Face Alignment and Reconstruction

TL;DR

CSANet addresses robust 3D face alignment and reconstruction from single 2D images, focusing on occlusion and lighting challenges. It augments a lightweight bottleneck backbone with Spatial Group-wise Enhancement and Coordinate Attention to refine features, and uses a joint Wing Loss and WPDC objective to stabilize 3DMM parameter learning. The approach yields superior accuracy, especially under large poses, and demonstrates faster training compared to baselines like 3DDFA, with competitive reconstruction quality on AFLW/AFLW2000-3D. This work provides a practical, attention-based framework for efficient, robust 3D face modeling suitable for real-world applications.

Abstract

Our project proposes an end-to-end 3D face alignment and reconstruction network. The backbone of our model is built by Bottle-Neck structure via Depth-wise Separable Convolution. We integrate Coordinate Attention mechanism and Spatial Group-wise Enhancement to extract more representative features. For more stable training process and better convergence, we jointly use Wing loss and the Weighted Parameter Distance Cost to learn parameters for 3D Morphable model and 3D vertices. Our proposed model outperforms all baseline models both quantitatively and qualitatively.
Paper Structure (22 sections, 10 equations, 6 figures, 2 tables)

This paper contains 22 sections, 10 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Face Alignment
  4. Face Reconstruction
  5. Attention
  6. Baseline Implementation
  7. 3D Morphable Model
  8. Network Structure and Cost Function
  9. Implementation Results
  10. Channel Spatial Attention Network
  11. Model Structure
  12. Attention Modules
  13. Spatial Group-wise Enhance Module
  14. Coordinate Attention Module
  15. Loss Function
  16. ...and 7 more sections

Figures (6)

  • Figure 1: 3DDFA Pipeline
  • Figure 2: The proposed model structure
  • Figure 3: Spatial Group-wise Enhance Module(SGE)
  • Figure 4: Coordinate Attention Module(CA)
  • Figure 5: Qualitative Visual Results of 3DDFA(a), DAMDNet(b) and our proposed model(c) on AFLW2000-3D dataset. Face alignment landmarks are shown on RGB images, with a consistent face reconstruction result
  • ...and 1 more figures