Noise-Tolerant Paradigm for Training Face Recognition CNNs
This work addresses the challenge of training high-performance face recognition models with noisy data, which is a domain-specific incremental improvement for computer vision applications.
The paper tackles the problem of noisy labels in large-scale face recognition datasets by proposing a noise-tolerant training paradigm that weights samples based on their probability of being clean, derived from the angular margin loss distribution, and demonstrates its effectiveness in experiments.
Benefit from large-scale training datasets, deep Convolutional Neural Networks(CNNs) have achieved impressive results in face recognition(FR). However, tremendous scale of datasets inevitably lead to noisy data, which obviously reduce the performance of the trained CNN models. Kicking out wrong labels from large-scale FR datasets is still very expensive, although some cleaning approaches are proposed. According to the analysis of the whole process of training CNN models supervised by angular margin based loss(AM-Loss) functions, we find that the $θ$ distribution of training samples implicitly reflects their probability of being clean. Thus, we propose a novel training paradigm that employs the idea of weighting samples based on the above probability. Without any prior knowledge of noise, we can train high performance CNN models with large-scale FR datasets. Experiments demonstrate the effectiveness of our training paradigm. The codes are available at https://github.com/huangyangyu/NoiseFace.