Aligning Across Large Gaps in Time
This method addresses challenges in remote sensing and UAV localization by providing robust image alignment, though it appears incremental as it builds on existing neural network and ICLK techniques.
The paper tackles the problem of temporally-invariant image registration across large time gaps, such as day-night, seasonal changes, and decades, for outdoor scenes, achieving transferability from satellite to ground-level webcam data.
We present a method of temporally-invariant image registration for outdoor scenes, with invariance across time of day, across seasonal variations, and across decade-long periods, for low- and high-texture scenes. Our method can be useful for applications in remote sensing, GPS-denied UAV localization, 3D reconstruction, and many others. Our method leverages a recently proposed approach to image registration, where fully-convolutional neural networks are used to create feature maps which can be registered using the Inverse-Composition Lucas-Kanade algorithm (ICLK). We show that invariance that is learned from satellite imagery can be transferable to time-lapse data captured by webcams mounted on buildings near ground-level.