Extending Absolute Pose Regression to Multiple Scenes
Hunter Blanton, Connor Greenwell, Scott Workman, Nathan Jacobs
Abstract
Direct pose regression using deep convolutional neural networks has become a highly active research area. However, even with significant improvements in performance in recent years, the best performance comes from training distinct, scene-specific networks. We propose a novel architecture, Multi-Scene PoseNet (MSPN), that allows for a single network to be used on an arbitrary number of scenes with only a small scene-specific component. Using our approach, we achieve competitive performance for two bench-mark 6DOF datasets, Microsoft 7Scenes and Cambridge Landmarks, while reducing the total number of network parameters significantly. Additionally, we demonstrate that our trained model serves as a better initialization for fine-tuning on new scenes compared to the standard ImageNet initialization, converging to lower error solutions within only a few epochs.