GADA: Geometry-Aware Deformable Aggregation for Image-Based Gaussian Splatting

Gaussian Splatting has achieved significant improvements by incorporating warping-based techniques. However, such methods suffer from pixel-level inaccuracies due to uncertain geometry. This uncertainty leads to spatial misalignments in the warped images, which disrupt residual learning used in warping-based methods and fundamentally limit the gains of correction, particularly on thin structures and high-frequency details. Driven by our insight that useful visual cues are not lost but locally preserved under slight displacement, we propose Geometry-Aware Deformable Aggregation (GADA). This method introduces an iterative refinement module with deformable offsets to actively correct spatial misalignments and recover these displaced cues. Furthermore, to address the limitations of standard pipelines where visibility checks (i.e., thresholding) often discard valid pixels and multi-view warped image fusion relies on naive mean aggregation, our module is coupled with an implicit confidence weighting mechanism that selectively suppresses unreliable evidence. Consequently, our approach outperforms prior warping-based Gaussian Splatting, preserving high-frequency quality while achieving 2.13 times faster FPS.

Creating realistic 3D scenes from ordinary photos is important for virtual reality, digital twins, and interactive graphics, but current methods can lose fine details such as thin bars, leaves, grass, and reflections. A common approach tries to improve a 3D scene by borrowing colors from nearby camera views, but if the estimated geometry is slightly wrong, those borrowed pixels are placed in the wrong location. Existing systems often handle this by throwing away suspicious pixels, which can also remove useful visual information. Our method, GADA, is based on a simple idea: many of these pixels are not wrong, just slightly displaced. Instead of discarding them, GADA searches nearby locations, realigns the useful visual cues, and learns how much to trust each camera view. This helps the model recover sharper textures and small structures while avoiding unreliable evidence. Across standard benchmarks for novel-view synthesis, GADA improves rendering quality over prior warping-based Gaussian Splatting methods and runs at 47 frames per second, more than twice as fast as the IBGS baseline. These improvements make high-quality 3D reconstruction more practical for real-time applications such as immersive viewing, digital twins, and 3D content creation.