Pengfei Wang, Shihao Wang, Liyi Chen, Zhiyuan Ma, Guowen Zhang, Lei Zhang
DepthMaster is a unified framework for generalized metric depth estimation on both perspective and 360-degree panoramic images.
Existing monocular depth estimation methods are specialized for specific camera types (perspective or panoramic), making it difficult to generalize accurate metric depth across both. This is due to geometric differences between perspective and panoramic cameras and the scarcity of panoramic training data with metric annotations.
DepthMaster decomposes panoramic images into overlapping perspective patches, processing each with a standard Transformer backbone. It introduces a novel Correspondence Consistency Loss (CCL) and injects virtual projection cameras as geometric priors to seamlessly stitch patches without specialized operators, maintaining a standard architecture. It also leverages powerful metric priors from vast perspective datasets to circumvent data scarcity.
Trained on a mixed dataset containing only one panorama dataset, DepthMaster achieves state-of-the-art zero-shot performance on 13 diverse datasets, outperforming not only universal methods but also leading specialist models in both perspective and panoramic domains.