Face recognition performance deteriorates when face images are of very low quality. For low quality video sequences, however, more discriminative features can be obtained by aggregating the information in video frames. We propose a Multi-mode Aggregation Recurrent Network (MARN) for real-world low-quality video face recognition. Unlike existing recurrent networks (RNNs), MARN is robust against overfitting since it learns to aggregate pretrained embeddings. Compared with quality-aware

more »

... methods, MARN utilizes the video context and learns multiple attention vectors adaptively. Empirical results on three video face recognition datasets, IJB-S, YTF, and PaSC show that MARN significantly boosts the performance on the low quality video dataset while achieves comparable results on high quality video datasets.