BA-Net: Dense Bundle Adjustment Networks

July 2020

tl;dr: Feature-metric differentiable bundle adjustment.

Overall impression

BA-Net proposed to do BA on feature maps to avoid sensitivity to photometric changes in the input, while still leverages denser information than keypoints. It also uses a new formulation of depth prediction from a series of prototype depth maps (cf YOLACT). –> The compact depth representation by coefficient is very similar to CodeSLAM where an AutoEncoder is used to compress depth maps. This compact depth representation decreases the search space of LM algorithm.

Note that there is no PoseNet to predict ego motion. The output of the BA layer is the camera pose sequence and point cloud depths.

The idea of feature metric loss is further extended in Feature metric monodepth ECCV 2020.

DeepV2D is similar to BA-Net.

Background of BA

Key ideas

Technical details