Sphere-Depth: A Benchmark for Depth Estimation Methods with Varying Spherical Camera Orientations
Researchers have introduced Sphere-Depth, a new benchmark designed to evaluate the performance of monocular depth estimation models when applied to spherical images. This benchmark specifically addresses the challenges posed by unintentional camera pose variations and the geometric distortions inherent in equirectangular projections, which are common in 360° vision applications. Experiments using Sphere-Depth revealed that even models designed for spherical imagery experience significant performance drops when camera orientation changes, highlighting a critical area for improvement in robotic navigation and immersive scene understanding. AI
IMPACT New benchmark highlights robustness issues in depth estimation for 360° vision, potentially guiding future model development for robotics and AR/VR.