Evaluation of Whole-Image Descriptors for Metric Localization

Appearance-based localization attempts to recover the position (and orientation) of a camera based on the images that it captured and a previously stored collection of images. Recent advances in image representations extracted using convolutional neural networks for the task of place recognition have produced whole-image descriptors which are robust to imaging conditions, including small viewpoint changes. In previous work, we have used these descriptors to perform localization by performing descriptor interpolation to compare the appearance of the image that is currently captured with the expected appearance at a candidate location. In this work, we directly study the behaviour of recently developed whole-image descriptors for this application.

[1]  Timo Schairer,et al.  Learning to localize with Gaussian process regression on omnidirectional image data , 2010, 2010 IEEE/RSJ International Conference on Intelligent Robots and Systems.

[2]  Nicolai Petkov,et al.  Image-based localization using Gaussian processes , 2016, 2016 International Conference on Indoor Positioning and Indoor Navigation (IPIN).

[3]  Thomas Brox,et al.  Synthesizing the preferred inputs for neurons in neural networks via deep generator networks , 2016, NIPS.

[4]  Michael Milford,et al.  Deep learning features at scale for visual place recognition , 2017, 2017 IEEE International Conference on Robotics and Automation (ICRA).

[5]  Javier González,et al.  Appearance-invariant place recognition by discriminatively training a convolutional neural network , 2017, Pattern Recognit. Lett..

[6]  Tomás Pajdla,et al.  NetVLAD: CNN Architecture for Weakly Supervised Place Recognition , 2018, IEEE Transactions on Pattern Analysis and Machine Intelligence.