GC-MVSNet
Making geometric consistency an explicit training signal for multi-view stereo.
Our multi-view, multi-scale geometric consistency checking process. During training, the geometric consistency of the estimated depth map is explicitly modeled across multiple source views. This allows the model to more quickly and accurately learn about geometric consistency, allowing the trained model to produce better reconstructions during inference.
Most multi-view stereo networks supervise each view’s depth map on its own, leaving the network to discover cross-view geometric consistency only implicitly — if at all. GC-MVSNet makes that consistency an explicit part of training.
During learning, the model checks each pixel’s predicted depth against the geometry implied by several other views at once, across multiple scales, and penalizes predictions that disagree. Instead of hoping consistency emerges, the network is pushed toward depths that hold up everywhere. The result is faster convergence and sharper reconstructions, and the mechanism drops into existing MVS pipelines without redesigning them.
The journal extension, GC-MVSNet++, develops the idea further and studies where geometric supervision helps most.