AnySplat: Feed-Forward 3D Gaussian Splatting from Unconstrained Views (city-super.github.io)

AnySplat is a newly introduced feed-forward network designed for novel-view synthesis from uncalibrated image collections, representing a significant advancement in the realm of AI and machine learning. Unlike traditional neural-rendering methods that require precise camera poses and scene-specific optimization, AnySplat forecasts 3D Gaussian representations of scene geometry and appearance in a single forward pass. This innovative model can efficiently handle casually captured datasets without needing pose annotations, making it particularly useful for practical applications where precise camera information is often unavailable. The technical framework of AnySplat utilizes a transformer-based geometry encoder followed by three decoders that predict Gaussian parameters, depth maps, and camera poses. This approach leads to the construction of pixel-wise 3D Gaussians, which are subsequently voxelized for rendering images and depth maps. In extensive evaluations, AnySplat demonstrates comparable quality to pose-aware methods while outperforming other pose-free techniques, significantly reducing rendering time and making real-time synthesis feasible in unconstrained settings. With its ability to handle both sparse and dense inputs efficiently, AnySplat paves the way for more accessible novel-view synthesis across various applications, from object-centric scenes to large-scale environments.