NTIRE 2024 Challenge on HR Depth from Images of Specular and Transparent Surfaces

This paper reports on the NTIRE 2024 challenge on HR Depth From images of Specular and Transparent surfaces, held in conjunction with the New Trends in Image Restoration and Enhancement (NTIRE) workshop at CVPR 2024. This challenge aims to advance the research on depth estimation, specifically to address two of the main open issues in the field: high-resolution and non-Lambertian surfaces. The challenge proposes two tracks on stereo and single-image depth estimation, attracting about 120 registered participants. In the final testing stage, 2 and 8 participating teams submitted their models and fact sheets for the two tracks.

1. Introduction

Recovering the 3D structure of a scene directly from images has been one of the most studied topics in computer vision. Depth estimation represents the first step for this purpose and a cornerstone for higher-level applications such as augmented reality, autonomous or assisted driving, robotics, and more. Although a variety of custom, active sensors exists for this task – LiDARs, Radars, Time-of-Flight (ToF), just to name a few – approaches estimating depth from one or multiple color images have gained higher and higher popularity with the advent of deep learning. Despite the steady improvements we witnessed in the last decade, estimating depth in certain conditions remains an open challenge. In particular, we identify two as the main sources of trouble.