ContextualNet: Exploiting Contextual Information Using LSTMs to Improve Image-Based Localization

Convolutional Neural Networks (CNN) have successfully been utilized for localization using a single monocular image [1]. Most of the work to date has either focused on reducing the dimensionality of data for better learning of parameters during training or on developing different variations of CNN models to improve pose estimation. Many of the best performing works solely consider the content in a single image, while the context from historical images is ignored. In this paper, we propose a combined CNN-LSTM which is capable of incorporating contextual information from historical images to better estimate the current pose. Experimental results achieved using a dataset collected in an indoor office space improved the overall system results to 0. 8 m & 2. 5° at the third quartile of the cumulative distribution as compared with 1. 5 m & 3. 0° achieved by PoseNet [1]. Furthermore, we demonstrate how the temporal information exploited by the CNN-LSTM model assists in localizing the robot in situations where image content does not have sufficient features.

Authors

• Mitesh Patel
• Brendan Emery
• Yan-Ying Chen

Keywords

• Feature extraction
• Cameras
• Context modeling
• Computer vision
• Logic gates
• Neural networks
• Contextual Information
• Long Short-term Memory
• Neural Network
• Convolutional Neural Network
• Deep Neural Network
• Short-term Memory
• Single Image
• Convolutional Neural Network Model
• Third Quartile
• Pose Estimation
• Office Space
• Indoor Spaces
• Sampling Rate
• Hyperparameters
• Image Features
• Recurrent Neural Network
• RGB Images
• Position Error
• Current Image
• Convolutional Neural Network Layers
• Orientation Of The Robot
• Scale-invariant Feature Transform
• Simultaneous Localization And Mapping
• Robot Pose
• Weights Of Layer
• Robot Operating System
• Structure From Motion
• Median Error
• Temporal Coherence