RLSplat: Resource-Linear Budget Control for Feed-forward 3D Gaussian Splatting

Key points

• To further improve its practicality, recent budget-controllable FF3DGS methods allow the number of Gaussians to be adjusted at inference, so that the output meets a given storage limit or rendering-latency constraint.1
• We therefore propose RLSplat, a resource-linear budget-controllable FF3DGS framework whose representation size, inference latency, and peak memory all decrease linearly as the Gaussian budget is reduced.1
• Instead of pruning or selection, RLSplat learns a Gaussian allocation policy that distributes the budget across local scene regions, so that exactly the requested number of Gaussians is generated.1
• In zero-shot evaluation, RLSplat outperforms prior budget-controllable methods at every budget, surpassing the strongest baseline even with 4$\times$ fewer Gaussians, while running up to 4.1$\times$ faster with 9.6$\times$ less peak memory at low budgets.1

Sources (1)

• [1]RLSplat: Resource-Linear Budget Control for Feed-forward 3D Gaussian Splatting
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 06:30 AM

  Feed-forward 3D Gaussian Splatting (FF3DGS) reconstructs 3D scenes from images in a single forward pass, making it well-suited to applications that must operate within limited resource budgets.

  To further improve its practicality, recent budget-controllable FF3DGS methods allow the number of Gaussians to be adjusted at inference, so that the output meets a given storage limit or rendering-latency constraint.

Extractive summary: sentences quoted from the sources.