Squeezeposenet: Image Based Pose Regression With Small Convolutional Neural Networks for Real Time Uas Navigation

ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci.

Sattler

Pollefeys

et al. 2019

Self Cite

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.

Section: Related Workmentioning

confidence: 99%

Image-to-Image Translation for Enhanced Feature Matching, Image Retrieval and Visual Localization

ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci.

Sattler

Pollefeys

et al. 2019

Self Cite

“…For the demands on pose regression SqueezePoseNet [10], an adaptation of SqueezeNet [31] is utilized. Whereas SqueezeNet is designed for solving classification tasks, SqueezePoseNet is modified to solve for pose regression.…”

Section: Modifications Of Cnnsmentioning

confidence: 99%

“…For the demands on UAV navigation, the utilization of a small CNN that can be used on-board computers is mandatory. Therefore, a CNN-based solution for the navigation or localization of an UAV in a known environment is utilized by SqueezePoseNet [10].…”

Section: Introductionmentioning

confidence: 99%

UAS Navigation with SqueezePoseNet—Accuracy Boosting for Pose Regression by Data Augmentation

Jutzi

2018

Drones

Abstract:The navigation of Unmanned Aerial Vehicles (UAVs) nowadays is mostly based on Global Navigation Satellite Systems (GNSSs). Drawbacks of satellite-based navigation are failures caused by occlusions or multi-path interferences. Therefore, alternative methods have been developed in recent years. Visual navigation methods such as Visual Odometry (VO) or visual Simultaneous Localization and Mapping (SLAM) aid global navigation solutions by closing trajectory gaps or performing loop closures. However, if the trajectory estimation is interrupted or not available, a re-localization is mandatory. Furthermore, the latest research has shown promising results on pose regression in 6 Degrees of Freedom (DoF) based on Convolutional Neural Networks (CNNs). Additionally, existing navigation methods can benefit from these networks. In this article, a method for GNSS-free and fast image-based pose regression by utilizing a small Convolutional Neural Network is presented. Therefore, a small CNN (SqueezePoseNet) is utilized, transfer learning is applied and the network is tuned for pose regression. Furthermore, recent drawbacks are overcome by applying data augmentation on a training dataset utilizing simulated images. Experiments with small CNNs show promising results for GNSS-free and fast localization compared to larger networks. By training a CNN with an extended data set including simulated images, the accuracy on pose regression is improved up to 61.7% for position and up to 76.0% for rotation compared to training on a standard not-augmented data set.

“…Therefore, pose estimation with CNNs is limited by the coverage and possible lack of training data. It was shown that pose regression in areas with less training data scores worse compared to areas with a dense distribution of training samples (Mueller et al, 2017). Utilizing a photorealistic model for data augmentation showed improvements regarding estimation accuracy (Mueller and Jutzi, 2018).…”

Section: Related Workmentioning

confidence: 99%

CNN-Based Initial Localization Improved by Data Augmentation

ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci.

Metzger

Jutzi

2018

Self Cite

Abstract. Image-based localization or camera re-localization is a fundamental task in computer vision and mandatory in the fields of navigation for robotics and autonomous driving or for virtual and augmented reality. Such image pose regression in 6 Degrees of Freedom (DoF) is recently solved by Convolutional Neural Networks (CNNs). However, already well-established methods based on feature matching still score higher accuracies so far. Therefore, we want to investigate how data augmentation could further improve CNN-based pose regression. Data augmentation is a valuable technique to boost performance on training based methods and wide spread in the computer vision community. Our aim in this paper is to show the benefit of data augmentation for pose regression by CNNs. For this purpose images are rendered from a 3D model of the actual test environment. This model again is generated by the original training data set, whereas no additional information nor data is required. Furthermore we introduce different training sets composed of rendered and real images. It is shown that the enhanced training of CNNs by utilizing 3D models of the environment improves the image localization accuracy. The accuracy of pose regression could be improved up to 69.37% for the position component and 61.61% for the rotation component on our investigated data set.