While existing multi-view stereoscopic (MVS) methods primarily rely on photometric and geometric consistency constraints, modern learning-based algorithms often rely on planar sweep algorithms to infer 3D geometry and apply explicit geometric consistency (GC) checks only in a postprocessing step, leaving the learning process itself unaffected. 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 learning process (see Figure 1). This integrated GC check directly penalizes geometrically inconsistent pixels, significantly accelerating the learning process and reducing the number of training iterations by half compared to other MVS methods. Furthermore, we 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 enforce multi-view, multi-scale supervised geometric consistency during training. The code is open source.
