Master thesis : Feed-Forward Novel View Synthesis For Soccer Scenes With Priors

Abstract

Novel View Synthesis ( NVS ) for soccer broadcasting would enable new production effects such as free-viewpoint replays, 3D camera transitions, immersive 3D analyses, etc. State-of-the-art NVS pipelines based on Neural Radiance Field (NeRF) and 3D Gaussian Splatting ( 3DGS ) produce high-quality novel views. However, their per-scene optimization, dense input requirements, and dependence on Structure-From- Motion ( SFM ) make them incompatible with the time and sparse input constraints of soccer broadcasting. Recent feed-forward 3D foundation models such as π3, Visual Geometry Grounded Transformer ( VGGT), and Depth-Anything-3 (DA3) recon- struct an entire scene in a single forward pass, but were neither trained nor evaluated on soccer scenes, where the scenes are large and the priors from the broadcast pipeline (camera calibration, pitch geometry, segmentation masks) are not used. Our work answers what a sparse-input feed-forward NVS pipeline made for soccer scenes would look like. We propose a modular four-stage architecture. A frozen 3D foundation model that outputs a coarse point cloud from a sparse input, a per-view depth alignment that aligns this point cloud to the broadcast world frame using the pitch as a reference, a learnable Dense Prediction Transformer (DPT) Gaussian head that regresses 3D Gaussians from the features of the foundation model, and a 3DGS rasterizer that renders novel views in real time. The pipeline is trained and evaluated on a feed-forward training and validation split that we create from SoccerNet-NVS. We benchmark three multi-view 3D foundation models and one monocular baseline on a custom evaluation pipeline. π3 ranks first on five of six metrics and is selected as the backbone. Our depth alignment beats global transform methods while preserving the broadcast world frame. On top of this aligned geometry, the trained Gaussian head produces plausible feed-forward novel views on a test scene in around one second per scene on a single GPU. Our pipeline is several orders of magnitude faster than the per-scene optimization methods. While the experiments do not currently produce high-quality NVS results, our approach presents a credible path towards real-time, photo-realistic 3D soccer scene reconstruction.