CF3: Compact and Fast 3D Feature Fields

Hyunjoon Lee, Joonkyu Min, Jaesik Park*
Seoul National University
* Corresponding author

Accepted to ICCV 2025 (Main Conference & Demonstrations Track)
arXiv Code

TL;DR: We propose a method to build compact and fast 3D Gaussian feature fields by effectively compressing and sparsifying Gaussians, achieving competitive performance with significantly fewer gaussians.

Teaser of CF3: Compact and Fast 3D Feature Fields


Abstract

3D Gaussian Splatting (3DGS) has begun incorporating rich information from 2D foundation models. However, most approaches rely on a bottom-up optimization process that treats raw 2D features as ground truth, incurring increased computational costs. We propose a top-down pipeline for constructing compact and fast 3D Gaussian feature fields, namely, CF3. We first perform a fast weighted fusion of multi-view 2D features with pre-trained Gaussians. This approach enables training a per-Gaussian autoencoder directly on the lifted features, instead of training autoencoders in the 2D domain. As a result, the autoencoder better aligns with the feature distribution. More importantly, we introduce an adaptive sparsification method that optimizes the Gaussian attributes of the feature field while pruning and merging the redundant Gaussians, constructing an efficient representation with preserved geometric details. Our approach achieves a competitive 3D feature field using as little as 5% of the Gaussians compared to Feature-3DGS.

Method

method overview
Overview of our CF3 pipeline. We utilize pre-trained 3D Gaussians to construct a 3D feature field. We adopt a weighted-sum strategy to lift features extracted from a visual foundation model into 3D. Subsequently, a per-Gaussian autoencoder compresses high-dimensional features into lower-dimensional embeddings, effectively removing noisy features through a variance filtering step. Afterward, adaptive sparsification merges redundant Gaussians, efficiently reducing the total Gaussian count and resulting in a compact 3D feature field.
The primary contributions of this work are as follows:
  1. We build a compact 3D feature field by lifting features via a pre-trained 3DGS and compressing them with a per-Gaussian autoencoder. This ensures robustness across downstream tasks since each Gaussian directly encodes view-consistent reference features.
  2. Our adaptive sparsification step optimizes the Gaussian feature field even further, which involves pruning and merging redundant Gaussians, while preserving essential details. As a result, our method achieves competitive performance while using as little as 5% of the original number of Gaussians, improving storage efficiency and rendering speed.

Adaptive Sparsification


Figure 2. Unlike the original 3D Gaussian Splatting, which preserves fine-grained details for photorealistic rendering, our method focuses on feature field reconstruction and merges redundant Gaussians to reduce unnecessary density, achieving effective sparsification.

LERF Dataset Results

SAM+CLIP Features


*Note that LangSplat uses three-levels mask(Whole, Part, SubPart), whereas we evaluate using a fixed mask level. We visualize results using the Whole mask.

Replica Dataset Results

MaskCLIP Features


Large Scale Dataset (KITTI) Results

SAM+CLIP Features


Custom Dataset Results

SAM+CLIP Features




BibTeX

@inproceedings{lee2025cf3,
  title={CF3: Compact and Fast 3D Feature Fields},
  author={Lee, Hyunjoon and Min, Joonkyu and Park, Jaesik},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={27906--27916},
  year={2025}
}

Reference

[1] Zhou, Shijie, et al. "Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.
[2] Qin, Minghan, et al. "Langsplat: 3d language gaussian splatting." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.
[3] Zhou, Chong, Chen Change Loy, and Bo Dai. "Extract free dense labels from clip." European conference on computer vision. Cham: Springer Nature Switzerland, 2022.