# FAST-Splat: Fast, Ambiguity-Free Semantics Transfer in Gaussian Splatting

**Authors:** Ola Shorinwa, Jiankai Sun, Mac Schwager  
**arXiv:** [2411.13753](https://arxiv.org/abs/2411.13753)  
**Date:** 2024-11-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2411.13753v1/

## Key points

- FAST-Splat introduced as a solution to limitations of existing semantic Gaussian Splatting methods
- Formulates open-vocabulary semantic Gaussian Splatting as an extension of closed-set semantic distillation
- Leverages explicit form of Gaussian Splatting scene representation for fast training and rendering speeds
- Augments each Gaussian with specific semantic codes instead of distilling semantics into separate neural fields
- Measures semantic similarity with open-vocabulary prompts and provides unambiguous semantic object labels and 3D masks
- Outperforms existing methods in terms of training speed, data pre-processing time, rendering speeds, and GPU memory usage
- Maintains similar or better semantic segmentation performance despite speed improvements

## Summary

, , , , 

In their paper titled "FAST-Splat: Fast, Ambiguity-Free Semantics Transfer in Gaussian Splatting," authors Ola Shorinwa, Jiankai Sun, and Mac Schwager introduce FAST-Splat as a solution to the limitations of existing semantic Gaussian Splatting methods. These limitations include slow training and rendering speeds, high memory usage, and ambiguous semantic object localization. <kw>FAST-Splat:</kw> addresses these challenges by formulating open-vocabulary semantic Gaussian Splatting as an extension of closed-set semantic distillation to the open-set setting. This enables precise semantic object localization results even when faced with ambiguous user-provided natural-language queries. By leveraging the explicit form of the Gaussian Splatting scene representation, FAST-Splat maintains the fast training and rendering speeds characteristic of Gaussian Splatting. Unlike previous methods that distill semantics into separate neural fields or utilize neural models for dimensionality reduction, FAST-Splat directly augments each Gaussian with specific semantic codes. This approach preserves the advantages of Gaussian Splatting in terms of training efficiency, rendering speed, and memory usage over neural field methods. The incorporation of Gaussian-specific semantic codes along with a hash-table allows FAST-Splat to measure semantic similarity with open-vocabulary prompts and provide unambiguous semantic object labels and 3D masks. Experimental results demonstrate that FAST-Splat outperforms existing methods by being 4x to 6x faster to train with a 13x faster data pre-processing step, achieving between 18x to 75x faster rendering speeds, and requiring approximately 3x smaller GPU memory compared to competing <kw>Gaussian Splatting</kw> methods. Despite these speed improvements, FAST-Splat maintains similar or better semantic segmentation performance. The authors plan to share links to the project website and codebase after the review period. Overall, <kw>FAST-Splat</kw> offers a significant advancement in fast and ambiguity-free semantics transfer within the realm of <kw>Gaussian Splatting</kw> techniques.