LEARNING ROBUST 3D FACE ALIGNMENT ARCHITECTURES WITH MONAS: A ONE-SHOT NAS AP

Abstract

3D face alignment remains a fundamental challenge in computer vision, particularly under real-world conditions such as extreme poses, occlusions, and lighting variations. Traditional deep learning methods often rely on manually designed architectures to regress 3D Morphable Model (3DMM) parameters or localize dense facial landmarks, requiring substantial domain expertise and exhibiting poor cross-domain generalization. To address these limitations, proposed MONAS (Multi-path One-shot Neural Architecture Search), a novel framework that autonomously discovers robust architectures for 3D face alignment. MONAS integrates multi-path feature extraction and contextual aggregation to effectively manage pose and scale variations. To enhance adaptability across diverse domains—including thermal images, cartoon faces, and medical scans—we embed domain adaptation modules into the architecture search space. MONAS introduces two key innovations: (1) Unbiased Multi-path Training to prevent path collapse during optimization, and (2) a Simulated Annealing-based NAS strategy for efficient and diverse architecture exploration. Extensive evaluations on standard and cross- domain benchmarks demonstrate that MONAS significantly outperforms existing handcrafted models in both sparse and dense landmark localization tasks