Rough or Noisy? Metrics for Noise Estimation in SfM Reconstructions

Abstract

Structure from Motion (SfM) can produce highly detailed 3D reconstructions, but distinguishing real surface roughness from reconstruction noise and geometric inaccuracies has always been a difficult problem to solve. Existing SfM commercial solutions achieve noise removal by a combination of aggressive global smoothing and the reconstructed texture for smaller details, which is a subpar solution when the results are used for surface inspection. Other noise estimation and removal algorithms do not take advantage of all the additional data connected with SfM. We propose a number of geometrical and statistical metrics for noise assessment, based on both the reconstructed object and the capturing camera setup. We test the correlation of each of the metrics to the presence of noise on reconstructed surfaces and demonstrate that classical supervised learning methods, trained with these metrics can be used to distinguish between noise and roughness with an accuracy above 85%, with an additional 5–6% performance coming from the capturing setup metrics. Our proposed solution can easily be integrated into existing SfM workflows as it does not require more image data or additional sensors. Finally, as part of the testing we create an image dataset for SfM from a number of objects with varying shapes and sizes, which are available online together with ground truth annotations.

Innovations in Photogrammetry and Remote Sensing: Modern Sensors, New Processing Strategies and Frontiers in Applications

“…Metrics were introduced and considered to present qualitative and quantitative assessments of the results. Finally, Nikolov and Madsen [ 5 ] introduced a metric for noise estimation in highly detailed 3D SfM reconstructions. In particular, the authors discussed a possible approach to distinguishing real surface roughness from reconstruction noise and proposed a number of geometrical and statistical metrics for noise assessment based on both the reconstructed object and the capturing camera setup.…”