Booster: A Benchmark for Depth From Images of Specular and Transparent Surfaces

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Current depth estimation methods excel but struggle with non-Lambertian materials and high-resolution image processing.
• Challenges remain in achieving accurate depth maps for specular, transparent, and high-resolution scenes.

Purpose of the Study:

• Introduce a novel dataset with high-resolution ground-truth labels for depth estimation.
• Address limitations in existing datasets regarding material properties and image resolution.
• Facilitate research in stereo and monocular depth estimation networks.

Main Methods:

• Developed a novel deep space-time stereo framework for data acquisition.
• Created a dataset with 606 high-resolution samples (12 Mpx) and unbalanced stereo pairs.
• Included manual material segmentation masks and unlabeled samples.

Main Results:

• The dataset features scenes with specular and transparent surfaces, crucial for challenging depth estimation.
• Provides sub-pixel precision ground-truth labels for accurate network evaluation.
• Experiments highlight existing challenges and future research directions in depth estimation.

Conclusions:

• The proposed dataset advances the field of depth estimation by providing challenging, high-resolution data.
• Enables better evaluation of stereo and monocular depth networks on complex materials.
• Opens new avenues for research in robust and accurate depth estimation.