ADODN: Attentive Dropout-Based Occlusion-Aware Deep Network for Facial Landmark Detection

Abstract

This paper introduces an attentive dropout-based occlusion-adaptive deep network (ADODN) for robust facial landmark detection under challenging conditions, including occlusions, extreme poses, and illumination variations. While convolutional neural networks have achieved high accuracy in Facial Landmark Detection (FLD), their performance often degrades under partial occlusions due to over-reliance on a few highly discriminative facial regions. ADODN addresses this limitation through three complementary modules: 1) a geometry-aware module that captures spatial relationships and structural priors among facial components; 2) an attentive dropout module that stochastically alternates drop masks and importance maps to encourage balanced feature learning from both dominant and subtle facial cues; and 3) a low-rank learning module that regularizes the regression representation by exploiting inter-feature correlations to recover occlusion-missing information. Unlike deterministic reweighting schemes, the attentive dropout mechanism improves robustness by randomly suppressing prominent responses during training, which mitigates feature over-dependence and promotes holistic structural inference. The resulting framework remains end-to-end and does not require auxiliary classifiers or multi-stage training. Extensive evaluations on challenging benchmarks (300W, COFW) show that ADODN achieves competitive and consistent performance, especially under occlusion-heavy settings. For example, ADODN attains 2.80 NRMSE on the 300W Common set and 5.81 on the Challenging set, improving upon recent baselines including ODN, AODN, and RHT-R. We also report parameter efficiency relative to prior ODN-style designs, supporting efficient training and inference.