Attention-Based Accurate and Robust Facial Landmark Detector

Recently, most of the existing face landmark de-tection algorithms focus on improving the performance and the efficiency of the model. However, due to the scaling and weighting effects as well as the existence of maximum pooling, the traditional convolutional neural network-based face alignment algorithms usually discard position and direction information implicit in the face image, thus decreasing its ability to capture the spatial position relationship between features, and resulting in significant performance degradation when partial occlusion occurs. To overcome this problem, in this paper, we propose an attention-based accurate, and robust face landmark detector called SCPNet. First, a SEBottleneck is used as the proposed detector's foundation module to explicitly describe channel in-terdependencies by adaptively recalibrating feature responses in terms of channels. The attention maps are then inferred sequentially along two independent dimensions (channel and space) by embedding a convolutional block attention module (CBAM) at the backbone network's tail, and the attention maps are multiplied by the input feature maps for adaptive feature refinement. Experimental results on WFLW, COFW, and 300W datasets show that our SCPNet outperforms current mainstream face alignment algorithms.