Learning Neighborhood-Reasoning Label Distribution (NRLD) for Facial Age Estimation

In this paper, we propose to learn a neighborhood-reasoning label distribution (NRLD) for facial age estimation. Unlike conventional label distribution methods with fixed-structural aging patterns, in this work, our NRLD aims to reason about more resilient and adaptive label distribution by disentangling the graph of face neighbors. In particular, our model holds the assumption on that the sample-specific age label distribution is principally influenced by a mixture of interpretable and meaningful factors, which typically cause plausible edges connected to the anchors. Under the scenario of each factor, we specifically collect the subset of graph edges and then convolute them with face samples to regress a mean-variance label distribution. During the training process, the mixture hyperparameters of our label distribution are iteratively optimized by following the Expectation-Maximization schema. Extensive experimental results on three challenging widely-evaluated datasets indicate the superiority in comparisons with most state of the arts.