BA-Net: Dense Bundle Adjustment Network

This paper introduces a neural network to solve the structure-from-motion (SfM) problem via feature bundle adjustment (BA), which explicitly enforces multi-view geometry constraints in the form of feature reprojection error. The whole pipeline is differentiable, so that the network can learn suitable feature representations that make the BA problem more trackable. Furthermore, this work introduces a novel depth parameterization to recover dense per-pixel depth. The network first generates some bases depth maps according to the input image, and optimizes the final depth as a linear combination of these bases via feature BA. The bases depth map generator is also learned via end-to-end training. The whole system nicely combines domain knowledge (i.e. hard-coded multi-view geometry constraints) and machine learning (i.e. feature learning and basis depth map generator learning) to address the challenging SfM problem. Experiments on large scale real data prove the success of the proposed method.