Fig. 3
Improvements to learning-based SLAM: Example image pairs (a) used to compute the reprojection error during training. From the two overlapping images \(I_1,I_2\), a respective weight mask (b) is applied to the re-projections (c) to calculate the re-projection error (d). This reduces the influence of unwanted classes on the training error (e), increasing focus on learning the geometry of coral reefs. Furthermore, the refraction-induced distortion of frames coming out of the camera (f) is rectified using the EUCM with the learned distortion parameters (g).
