Edges Compete for Trust: Group Relative Edge Optimization for Building Reconstruction from Point Clouds

Confidence >> (b) GREO for Edge-based Methods Rewards >> >> >> Edge Proposals Ground Truth group advantages low high Sparse Supervision Input & GT EdgeDiff EdgeDiff+GREO (c) Reconstruction Results Dense Supervision Figure 1. Comparison between previous methods and our GREO. (a) Previous edge-based methods rely on sparse supervision via Hungarian matching, where only a small subset of matched edges receive effective gradients. (b) Our GREO introduces dense supervision by computing group-relative advantages for all edge proposals, enabling discriminative confidence optimization. (c) GREO integrates seamlessly into existing edge-based methods, introducing no inference overhead while outperforming the state-of-the-art EdgeDiff [25].