MAFL (Multi-Attribute Facial Landmark)

人脸对齐的人脸地标检测长期以来一直受到遮挡和姿态变化问题的阻碍。我们没有将检测任务视为单个独立的问题，而是研究了通过多任务学习提高检测鲁棒性的可能性。具体来说，我们希望优化面部标志检测以及异构但微妙相关的任务，例如头部姿势估计和面部属性推断。这是非常重要的，因为不同的任务具有不同的学习难度和收敛速度。为了解决这个问题，我们制定了一种新的任务约束深度模型，具有任务明智的提前停止以促进学习收敛。广泛的评估表明，所提出的任务约束学习 (i) 优于现有方法，尤其是在处理具有严重遮挡和姿势变化的面部时，并且 (ii) 与基于级联深度模型。

Facial landmark detection for face alignment has long been hampered by occlusion and pose variation issues. Rather than treating the detection task as an isolated standalone problem, we investigate the feasibility of enhancing detection robustness via multi-task learning. Specifically, we aim to optimize facial landmark detection alongside heterogeneous but subtly correlated tasks such as head pose estimation and facial attribute inference. This is highly important as different tasks have varying learning difficulties and convergence speeds. To address this problem, we propose a novel task-constrained deep model with task-wise early stopping to promote learning convergence. Extensive evaluations show that the proposed task-constrained learning (i) outperforms existing methods, especially when handling faces with severe occlusion and pose variations, and (ii) compares favorably with cascaded deep models.

搜集汇总

背景与挑战

背景概述

MAFL是一个专注于多任务学习的人脸地标检测数据集，特别针对面部遮挡和姿态变化问题设计，由香港中文大学团队开发并于2014年发布。该数据集通过结合头部姿势估计和面部属性推断等辅助任务，显著提高了在复杂条件下的人脸特征检测性能。