Abstract
Abstract. The performance of machine learning and deep learning algorithms for image analysis depends significantly on the quantity and quality of the training data. The generation of annotated training data is often costly, time-consuming and laborious. Data augmentation is a powerful option to overcome these drawbacks. Therefore, we augment training data by rendering images with arbitrary poses from 3D models to increase the quantity of training images. These training images usually show artifacts and are of limited use for advanced image analysis. Therefore, we propose to use image-to-image translation to transform images from a rendered domain to a captured domain. We show that translated images in the captured domain are of higher quality than the rendered images. Moreover, we demonstrate that image-to-image translation based on rendered 3D models enhances the performance of common computer vision tasks, namely feature matching, image retrieval and visual localization. The experimental results clearly show the enhancement on translated images over rendered images for all investigated tasks. In addition to this, we present the advantages utilizing translated images over exclusively captured images for visual localization.
Highlights
The performance of common machine learning algorithms typically scales with the quantity and quality of training data utilized to optimize them
Deep learning with Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) pushed the performance of learning based approaches in the recent years
(iii) we show that image-to-image translation concerning 3D models enhances performance of common computer vision tasks
Summary
The performance of common machine learning algorithms typically scales with the quantity and quality of training data utilized to optimize them. The demand for training data increased and training data sets for numerous tasks were recently published In this contribution, we generate new training images by image-to-image translation to subsequently improve performance of common computer vision and photogrammetry tasks. Augmenting training data is a powerful option to overcome challenges in several fields of computer vision, like feature matching, image retrieval and visual localization Such data augmentation includes the modification of existing training images as well as the generation of new images to expand training sets. Common methods in image processing are to shift, rotate, scale, flip, crop, transform, compress or blur training images to extend a basis data set In this contribution new images are rendered and translated by a GAN to augment a data set of images. If more variety of training samples is considered in a training set, more robust and accurate networks can be expected
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
