Robust Fitting with Truncated Least Squares: A Bilevel Optimization Approach
Huu Minh Le and Christopher Zach
We consider the problem of robust fitting with truncated least squares cost function. Existing approaches involve replacing the truncated least squares by a smooth approximation that allows the problem to be solved using variants of Iteratively Re-weighted Least Squares (ILRS). In this work, we propose a new approach based on bi-level optimization that leads to a new algorithm to compute residual weights for the truncated least squares loss, which enables us to incorporate our new approach to existing non-linear least squares solvers. Experimental results show promising results on several large-scale bundle adjustment instances.