Challenges of Depth Estimation for Transparent Objects

Abstract

Transparent objects and surfaces are pervasive in man-made environments and need to be considered in any vision system. Accurate depth data is a key factor for such systems reliability, requiring transparency to be inferred, due to the sensing challenges. However, the current state-of-the-art methods to predict the depth of such objects are not reliable enough to ensure safe operation of robots in arbitrary complex scenes. In order to better understand and improve upon existing solutions, we evaluate the performance of a variety of depth estimation methods. Doing so, we disentangle the different factors impacting their performance. Among our findings, neural radiance fields offer the best accuracy, but are very sensitive to the number of images used to understand the scene, and do not benefit from any level of object understanding to help them fill in the gaps.