[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.

Benchmark Datasets

In the papaer, we introduce two new challenging GFS-PCS benchmarks with diverse novel classes for comprehensive generalization evaluation.

This repository contains the two novel GFS-PCS benchmarks:

ScanNet200: Our GFS benchmark based on ScanNet200, also including the original ScanNet labels
ScanNet++: Our GFS benchmark based on ScanNet++

These benchmarks lay a solid foundation for real-world GFS-PCS advancements.

Note: To use these datasets, you must first agree to the terms and apply for access. Please refer to ScanNet200 and ScanNet++ for instructions.

Dataset Structure

Each dataset in this repository is organized as follows:

Splits: Train, validation, and test sets.
Registration Data List: For both 1-shot and 5-shot scenarios, each includes five randomly generated registration sets.

Usage

For detailed usage instructions, model implementation, and training code, please refer to our GitHub repository.

Pre-trained Model Weights

The complete GFS-VL pre-trained model weights can be found in our model weights 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}
}