Early Adaptation of Deep Priors in Age Prediction From Face Images

Mahdi Hajibabaei, Anna Volokitin, Radu Timofte; Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 1639-1647

Age prediction from face images is a challenging task. Direct application of pre-trained models on new data leads to poor performance due to data and distribution mismatch and lack of newly annotated material. In this work, we analyze the transfer of knowledge from deep models pre-trained on massive datasets to new target datasets with (very) little information available. We investigate (i) pre-training on massive datasets with an imposed target age label distribution, (ii) pre-training on massive face datasets but without age annotations, and (iii) fine-tuning on the target train data. The experimental benchmark uses the massive IMDB-Wiki, VGG-Face and ImageNet datasets as sources and ChaLearn LAP and MORPH 2 as target datasets. The deep architectures/priors are based on the VGG-16 and the recent state-of-the-art DEX and VGG-Face models. Our main findings are as follows. (i) Using deep priors (pre-trained models on similar data and/or task) boosts the performance on the target dataset. (ii) Imposing the target age label distribution on pre-trained models helps. (iii) The access to and the use of labeled target samples is critical - with as few as 12 samples used for fine-tuning a large performance gain is achieved, surpassing the impact of imposing target distribution for pre-training. Early adaptation of deep priors to new target datasets can yield sufficiently good performance at a reasonably low computational cost.