While existing multi-view stereoscopic (MVS) methods primarily rely on photometric and geometric consistency constraints, recent learning-based algorithms rely on planar sweep algorithms to infer 3D geometry, and explicit geometric consistency (GC) checks are applied only in the postprocessing stage and do not affect the training process itself. In this study, we present GC MVSNet plus plus, a novel method that actively enforces geometric consistency of reference view depth maps across multiple source views (multi-view) and scales (multi-scale) during the training stage. This integrated GC check significantly accelerates the training process by directly penalizing geometrically inconsistent pixels, reducing the number of training iterations by half compared to other MVS methods. We also present a densely connected cost regularization network with two unique block designs (simple and feature-dense) optimized to leverage dense feature connections for improved regularization. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance on the DTU and BlendedMVS datasets and ranks second on the Tanks and Temples benchmark. GC MVSNet plus plus is the first method to enhance supervised geometric consistency during training across multiple views and multiple scales. The code is publicly available.
