Enhancing Matching Response Discriminability and Complementary Multi‐Scale Cost Aggregation for Unsupervised Multi‐View Stereo

ABSTRACT

Feature extraction and cost volume regularisation are crucial in learning‐based Multi‐view Stereo (MVS). However, existing unsupervised MVS approaches often focus primarily on constructing supervision signals, paying less attention to feature matching discriminability and receptive fields during cost volume construction and regularisation. To address this, we propose two modules, InceptionMamba and InceptionConv3D, to replace standard 2D and 3D convolutions in unsupervised MVS networks respectively. Inspired by multi‐perspective human observation, InceptionMamba employs four parallel branches to handle varying surrounding pixels around matching points caused by viewpoint changes, enhancing feature discriminability via multi‐scale features. InceptionConv3D aggregates multi‐scale features along each depth plane of the cost volume, providing advantageous cues for reference‐view depth estimation. Furthermore, we directly apply the two modules to CL‐MVSNet, a state‐of‐the‐art unsupervised end‐to‐end method, and thereby propose INT‐MVSNet. Extensive experiments on DTU and Tanks & Temples benchmarks validate our approach. On DTU, compared to CL‐MVSNet, our method achieves improvements of 1.33%, 3.18%, and 2.13% in accuracy, completeness, and overall metrics, respectively. Additionally, on Tanks & Temples, our method achieves the second‐best overall performance among all compared methods.