An analysis of sample synthesis for deep learning based object detection

• Ano de defesa: 2020
• Autor(a) principal: Leonardo Blanger
• Orientador(a): Nina Sumiko Tomita Hirata
• Banca de defesa: David Menotti Gomes, Roberto de Alencar Lotufo
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade de São Paulo
• Programa de Pós-Graduação: Ciência da Computação
• País: BR
• Link de acesso: https://doi.org/10.11606/D.45.2020.tde-10112020-203810

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.