2019/10/31 by Valentin Vielzeuf, Vielzeuf, Valentin, Alexis Lechervy +5 · 1 citation
Computer Science · #Face recognition and analysis #Domain Adaptation and Few-Shot Learning #Generative Adversarial Networks and Image Synthesis
paper · pdf · doi:10.48550/arxiv.1911.03222
This paper proposes a step toward obtaining general models of knowledge for\nfacial analysis, by addressing the question of multi-source transfer learning.\nMore precisely, the proposed approach consists in two successive training\nsteps: the first one consists in applying a combination operator to define a\ncommon embedding for the multiple sources materialized by different existing\ntrained models. The proposed operator relies on an auto-encoder, trained on a\nlarge dataset, efficient both in terms of compression ratio and transfer\nlearning performance. In a second step we exploit a distillation approach to\nobtain a lightweight student model mimicking the collection of the fused\nexisting models. This model outperforms its teacher on novel tasks, achieving\nresults on par with state-of-the-art methods on 15 facial analysis tasks (and\ndomains), at an affordable training cost. Moreover, this student has 75 times\nless parameters than the original teacher and can be applied to a variety of\nnovel face-related tasks.\n