Adaptive Keyframe Selection for Online Iterative NeRF Construction

Joshua Wilkinson, Jack Naylor, Ryan Griffiths, Donald G. Dansereau

Keywords: nerf, mapping, slam, persistent mapping

TL;DR: We propose an adaptive keyframe selection method that allows robots to iteratively build NeRF-based maps from full frame rate video and over extended deployments.

Abstract: We propose a method for intelligently selecting images for building neural radiance fields (NeRFs) from the large number of frames available in typical robot-mounted cameras. Our approach iteratively constructs and queries a NeRF to adaptively select informative frames. We demonstrate that our approach maintains high-quality representations with a 78\% reduction in input data and reduced training time in single-pass mapping, while preventing unbounded growth of input frames in persistent mapping. We also demonstrate our adaptive approach outperforming non-adaptive spatial and temporal methods in terms of training time and rendering quality. This work is a step towards persistent robotic NeRF-based mapping.

Submission Number: 19