Efficient Merging of Maps and Detection of Changes

Generic Merging of Structure from Motion Maps with a Low Memory Footprint

“…The proposed method exploits the fact that the optimal residuals from the separate bundles can be linearised to avoid the large bundles. Our bundle representation is built on theory from [21] and for completeness, we will summarise some of that theory in this section.…”

“…To summarise the theory from [21], the compressed representation of data consists of (q| o , a, R), where q| o is a subset of the 3D points. Note that while J q is a rectangular matrix, R will be quadratic and thus much smaller than J q .…”

“…Note that while J q is a rectangular matrix, R will be quadratic and thus much smaller than J q . Despite this, it was shown in [21] that it is possible to obtain a good approximation of the residual according to Equation (8). Furthermore, once an update has been made, the rest of the points and the camera matrices can be updated using ∂s/∂q.…”

“…In [21], a method that is a compromise between a full optimisation bundle and the Kalman filter was presented. The method was primarily evaluated on audio data together with a small experiment for image data.…”

“…In this paper we develop that idea further to work automatically for SfM data from RGB images. The idea behind the method presented in [21] is that maps can be merged efficiently using only a small memory footprint from the map and the residuals. Then the merging problem can be solved linearly.…”