Joint face detection and Facial Landmark Localization using graph match and pseudo label

Face detection and Facial Landmark Localization (FLL) are not integrated well because training samples with annotations of both bounding box and facial landmarks are costly to get. This paper presents a real-time framework for joint face detection and FLL. We exploit the synergy between the two tasks, then design a fully convolutional network to predict the location of facial landmarks and face regions. Besides, we observe the cluster assumption in FLL and thus propose a progressively pseudo labeling training that not only eliminates the harmful effect caused by inaccurate/noisy annotations, but also makes full use of exact, inexact and coarse-grained labels. To complete the bottom-up face assembly after model inference, we proposed two graph matching algorithms without learnable parameters. The whole framework keeps end-to-end attribute. Extensive experiments show that our approaches achieve state-of-the-art level on face detection and FLL in the different datasets, both in accuracy and runtime.

Automated summary

This paper proposes a real-time framework for joint face detection and Facial Landmark Localization (FLL). It utilizes a fully convolutional network to predict the location of facial landmarks and face regions, and introduces a progressively pseudo labeling training method to eliminate the effect of inaccurate/noisy annotations. Two graph matching algorithms without learnable parameters are also proposed for completing the bottom-up face assembly.