SEFD: A Simple and Effective Single Stage Face Detector

Recently, the state-of-the-art face detectors are extending a backbone network by adding more feature fusion and context extractor layers to localize multi-scale faces. Therefore, they are struggling to balance the computational efficiency and performance of face detectors. In this paper, we introduce a simple and effective face detector (SEFD). SEFD leverages a computationally light-weight Feature Aggregation Module (FAM) to achieve high computational efficiency of feature fusion and context enhancement. In addition, the aggregation loss is introduced to mitigate the imbalance of the power of feature representation for the classification and regression tasks due to the backbone network initialized by the pre-trained model that focuses on the classification task other than both the regression and classification tasks. SEFD achieves state-of-the-art performance on the UFDD dataset and mAPs of 95.3%, 94.1%, 88.3% and 94.9%, 94.0%, 88.2% on the easy, medium and hard subsets of WIDER Face validation and testing datasets, respectively.