Deeply Learned Rich Coding for Cross-Dataset Facial Age Estimation

We propose a method for leveraging publicly available labeled facial age datasets to estimate age from unconstrained face images at the ChaLearn Looking at People (LAP) challenge 2015 [9]. We first learn discriminative age related representation on multiple publicly available age datasets using deep Convolutional Neural Networks (CNN). Training CNN is supervised by rich binary codes, and thus modeled as a multi-label classification problem. The codes represent different age group partitions at multiple granularities, and also gender information. We then train a regressor from deep representation to age on the small training dataset provided by LAP organizer by fusing random forest and quadratic regression with local adjustment. Finally, we evaluate the proposed method on the provided testing data. It obtains the performance of 0.287, and ranks the 3rd place in the challenge. The experimental results demonstrate that the proposed deep representation is insensitive to cross-dataset bias, and thus generalizable to new datasets collected from other sources.