DepthAgent: Towards Better Universal Depth Estimation via Sample-wise Expert Selection
Researchers have developed DepthAgent, a novel vision-language agent designed to improve monocular depth estimation across various camera types. Unlike previous methods that use a single estimator, DepthAgent leverages multiple pre-existing depth models as tools. It intelligently analyzes scene and camera geometry to select or fuse predictions from these experts, particularly excelling on challenging samples where individual models falter. This adaptive approach significantly enhances accuracy and robustness for depth estimation tasks. AI
IMPACT Enhances depth estimation accuracy and robustness across diverse camera geometries by adaptively selecting and fusing multiple expert models.