[1/4] Camera poses are essential for (neural) 3D reconstruction. But what about sparse-view settings where obtaining these via COLMAP isnโt feasible? Our ECCV paper tackles this using an energy-based formulation for predicting relative rotation (jasonyzhang.com)
[2/N] We use the pairwise predictions to find a maximally consistent set of rotations across multiple images. Starting with a greedy initialization, we iteratively update rotations to maximize the joint likelihood.
![[2/N] We use the pairwise predictions to find a maximally consistent set of rotations across multipl...](https://pbs.twimg.com/media/Fa3dC1BXwAAiZf0.jpg)
[3/N] This approach is more robust in sparse-view settings compared to SfM-based methods (although you should still use COLMAP given many views!), and also generalizes to unseen object categories.
![[3/N] This approach is more robust in sparse-view settings compared to SfM-based methods (although y...](https://pbs.twimg.com/media/Fa3dJrNWAAIvB2l.jpg)
[4/N] Assuming centre-facing cameras, the recovered rotations can initialize 6D poses and help bootstrap sparse-view neural 3D reconstruction methods (e.g. @jasonyzhang2's prior work on NeRS).
![[4/N] Assuming centre-facing cameras, the recovered rotations can initialize 6D poses and help boots...](https://pbs.twimg.com/media/Fa3d_AmXwAIUKvM.jpg)
Loading suggestions...