3D Gaussian Splatting (3D-GS) is a recent scene reconstruction technique that enables real-time rendering of new scene views. It accomplishes this by learning scene representations modeled as parametric point clouds of 3D Gaussians. However, this technique faces inefficiencies in model size and rendering speed, particularly in resource-constrained environments. This proposal introduces two techniques to address these inefficiencies: PUP 3D-GS and Speedy-Splat.

PUP 3D-GS is a post-hoc pruning technique that computes the importance of each Gaussian to the reconstructed scene as a per-Gaussian score. It then iteratively removes the least important Gaussians and fine-tunes the scene. High image quality is maintained even after removing 90% of Gaussians.

Speedy-Splat integrates PUP 3D-GS pruning into the training process by reparameterizing the importance score while also enhancing Gaussian localization during rendering. These changes boost rendering speeds by 6.71x across real-world scenes from three popular datasets.

In addition to these advancements, this proposal discusses two works in AI for social good that have received positive recognition from the research community. The first demonstrates that training data with heavy racial bias can still produce surprisingly fair inference. The second introduces a novel architecture designed to detect violence in videos.