CVPR Daily - Wednesday

popularity and found applications in various domains, issues have emerged when dealing with many different outliers. While previous methods might perform well with 100 people in the scene, they fail when confronted with diverse shapes like a dog, ball, or box. In contrast, this paper’s optimization-based approach treats all outliers equally , regardless of their shape. “ At first, we assumed that by using a robust loss and treating it like a typical robust optimization, where the loss takes care of the outliers, it would work fine, but in practice, it wasn’t working, ” Sara reveals. “ If you do it yourself by replacing the reconstruction loss with a robust loss, it will fail because of the trajectory of the optimization that NeRF follows. At the start of the training, the details of the scene, which makes NeRF very good, will always have a high loss. They will be trimmed if you just blindly use a robust loss. Your images will become blurred and without any interesting details. Our main struggle was balancing this and separating the transient objects as outliers , as opposed to early training details or viewpoint effects. ” The model can also handle glossy or transparent objects . It does not remove view-dependent artifacts that are intentional, only transient objects, because of a loss that considers the spatial consistency of an object. “ We use an iteratively reweighted least-squares (IRLS) with a Trimmed least squares (LS) loss , a specific robust loss that adapts the threshold 4 DAILY CVPR Wednesday Highlight Presentation