Illumination changes


Frames from sequences used for evaluating performance in the presence of illumination changes from the ETH Illumination dataset

Many SLAM algorithms assume that lighting will be consistent and rely on color information to estimate the camera position. In practice, lighting conditions often change unpredictably - for example if a light source is temporarily occluded. Here we study the effect of lighting changes on two dense systems which use a photometric error as part of their pose computation, ReFusion and ElasticFusion, and two sparse, feature-based algorithms: ORB-SLAM2 and ORB-SLAM3.

We use the ETH Illumination dataset, which is based on the TUM and ICL-NUIM datasets, containing both real and synthetic sequences with local and global illumination changes, as well as sequences where a flashlight shining from the perspective of the camera is used. It is immediately clear from the results that thanks to lighting-invariance, the feature-based algorithms perform significantly better.