Convolutional Neural Network-based Place Recognition

By Zetao Chen, Obadiah Lam, Adam Jacobson, Michael Milford. 2014

276 citations recorded by the source corpus.

Recently Convolutional Neural Networks (CNNs) have been shown to achieve state-of-the-art performance on various classification tasks. In this paper, we present for the first time a place recognition technique based on CNN models, by combining the powerful features learnt by CNNs with a spatial and sequential filter. Applying the system to a 70 km benchmark place recognition dataset we achieve a 75% increase in recall at 100% precision, significantly outperforming all previous state of the art techniques. We also conduct a comprehensive performance comparison of the utility of features from all 21 layers for place recognition, both for the benchmark dataset and for a second dataset with more significant viewpoint changes.

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition, Electrical and Electronic Engineering

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Aerospace Engineering

1. Michael Milford
2. Niko Sünderhauf
3. Lei Tai
4. Wolfram Burgard
5. Cyrill Stachniss
6. Ben Upcroft
7. John J. Leonard
8. Paul M. Newman

In Computer Vision and Pattern Recognition

1. Lei Tai
2. Luc Van Gool
3. Chunhua Shen
4. Michael Milford
5. Niko Sünderhauf
6. Vladlen Koltun
7. Andrew Zisserman
8. Andrea Vedaldi

In Electrical and Electronic Engineering

1. Lei Tai
2. Niko Sünderhauf
3. Michael Milford
4. Ingmar Kanitscheider
5. Ila Rani Fiete
6. John J. Leonard
7. Wolfram Burgard
8. Cyrill Stachniss

Citations in the corpus

• On the performance of ConvNet features for place recognition (2015, methodology): “[30] was the first to consider ConvNets for place recognition.”
• Place Recognition with ConvNet Landmarks: Viewpoint-Robust, Condition-Robust, Training-Free (2015, methodology): “We are investigating whether fine tuning the network for the specific task of place recognition will result in further performance gains for our approach, but also for methods such as [41, 3].”
• Deep learning features at scale for visual place recognition (2017, background): “Motivated by these results, recent studies have also shown that state-of-the-art performance in place recognition can be achieved by utilizing intermediate representations from CNNs that have already been trained on object recognition datasets [8-10].”
• Visual Place Recognition: A Survey (2016, methodology): “The authors of [206], [207] utilized CNN features as holistic image descriptors and analyzed the robustness of different layers against visual”
• Deep-learning in Mobile Robotics - from Perception to Control Systems: A Survey on Why and Why not (2016, methodology): “Actually, the convolutional filter is a typical image processing method and CNN can be re-JOURNAL garded as a deformation of the traditional mathematical model.”
• Training recurrent networks to generate hypotheses about how the brain solves hard navigation problems (2016, methodology): “[27, 28]) from camera images, but the geometric calculations and parameters at the core of SLAM algorithms are still largely hand-specified.”
• Robust visual SLAM across seasons (2015, methodology): “A very recent approach has used these features to perform visual place recognition [4], but does not address matching images across seasons, and assumes linear trajectories.”