[CVPR 2025] GFS-VL: Generalized Few-shot 3D Point Cloud Segmentation with Vision-Language Model
Overview
GFS-VL is a novel framework proposed in our CVPR 2025 paper: Generalized Few-shot 3D Point Cloud Segmentation with Vision-Language Model.
Our approach leverages the synergy between:
- Dense but noisy pseudo-labels from 3D Vision-Language Models
- Precise yet sparse few-shot samples
by maximizing the strengths of both data sources for effective generalized few-shot 3D point cloud segmentation.
Released Model Weights
This repository contains the following pre-trained weights:
- PTv3 Backbones: Our pre-trained point transformer v3 backbones
- GFS-VL Models: Complete GFS_VL few-shot segmentation framework
Usage
For detailed usage instructions, model implementation, and training code, please refer to our GitHub repository.
Benchmarks
We introduce two new challenging GFS-PCS benchmarks with diverse novel classes for comprehensive generalization evaluation. These benchmarks lay a solid foundation for real-world GFS-PCS advancements.
The benchmark datasets can be downloaded from our Huggingface dataset repository.
Citation
If you find our work useful, please consider citing our paper:
@inproceedings{an2025generalized,
title={Generalized Few-shot 3D Point Cloud Segmentation with Vision-Language Model},
author={An, Zhaochong and Sun, Guolei and Liu, Yun and Li, Runjia and Han, Junlin and Konukoglu, Ender and Belongie, Serge},
booktitle=CVPR,
year={2025}
}