Download como arquivo ICAL
"An Analysis of Sample Synthesis for Deep Learning based Object Detection"
Sexta-feira 16 Outubro 2020, 16:00
Contato: Este endereço de email está sendo protegido de spambots. Você precisa do JavaScript ativado para vê-lo.

Candidato: Leonardo Blanger

Orientadora: Profa. Dr. Nina Sumiko Tomita Hirata



BLANGER, L. An Analysis of Sample Synthesis for Deep Learning based Object Detection. 2020. Dissertation (Master) - Institute of Mathematics and Statistics, University of São Paulo, São Paulo, 2020.

This work investigates the use of artificially synthesized images as an attempt to reduce the dependency of modern Deep Learning based Object Detection techniques on expensive supervision. In particular, we propose using a big number of synthesized detection samples to pretrain Object Detection architectures before finetuning them on real detection data. As the major contribution of this project, we experimentally demonstrate how this pretraining works as a powerful initialization strategy, allowing the models to achieve competitive results using only a fraction of the original real labeled data. Additionally, in order to synthesize these samples, we propose a synthesis pipeline capable of generating an infinite stream of artificial images paired with bounding box annotations. We demonstrate how it is possible to design such a working synthesis pipeline just using already existing GAN techniques. Moreover, all stages in our synthesis pipeline can be fully trained using only classification images. Therefore, we managed to take advantage of bigger and cheaper classification datasets in order to improve results on the harder and more supervision hungry Object Detection problem. We demonstrate the effectiveness of this pretraining initialization strategy combined with the proposed synthesis pipeline, by performing detection using four real world objects: QR Codes, Faces, Birds and Cars.

Keywords: Object Detection, Sample Synthesis, Generative Models, Deep Learning.