Progressive Batching for Efficient Non-linear Least Squares

Non-linear least squares solvers are used across a broad range of offline and real-time model fitting problems. Most improvements the basic Gauss-Newton algorithm tackle convergence guarantees or leverage sparsity underlying problem structure for computational speedup. With success deep learning methods leveraging large datasets, stochastic optimization received recently lot attention. Our work borrows ideas from both machine statistics, we present an approach non-linear least-squares that while at same time significantly reduces required amount computation. Empirical results show our proposed method achieves competitive rates compared to traditional second-order approaches on common computer vision problems, such as image alignment essential matrix estimation, with very numbers residuals.