Institutional Repository of Vilnius University
Vaizdų transformacija augmentacijoms naudojantis giliaisiais neuroniniais tinklais /
Abstract
dc:descriptionTraining a convolutional neural network is the first step towards solving computer vision problems. To make this happen, an image dataset has to be used. In some cases, the dataset doesn't have enough data. To solve this problem, one can attempt to create data themselves, but that requires a lot of time and monetary investments. Another solution is to use image augmentations which can still be insufficient. In this case, more data can be created by utilizing image-to-image translations using generative adversarial networks. The goal of this work is to discover how object detection results are influenced by using translated images for the finetuning process. CycleGAN, DRIT and MSGAN were chosen as the models for image-to-image translation. Object detection results were measure after finetuning Faster-RCNN + ResNet101 and SSD + MobileNetV2 models. BDD100K and „Transient Attribute“ datasets were used during the finetuning process. The results show that image augmentations give only slightly better results than using translated images when measuring the object detection metrics. Another finding is that generative adversarial network LPIPS metric measurements can be used as an indicator for how GAN's translated images can positively influence the object detection model performance after finetuning.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Paškevičius, Saulius,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD146241870&prefLang=en_US