# TSP3D: Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding

This repo contains the models for paper [Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding](https://arxiv.org/abs/2502.10392). Code is available at: [https://github.com/GWxuan/TSP3D](https://github.com/GWxuan/TSP3D)

> Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding  
> [Wenxuan Guo](https://github.com/GWxuan)*, [Xiuwei Xu](https://xuxw98.github.io/)\*, [Ziwei Wang](https://ziweiwangthu.github.io/), [Jianjiang Feng](https://ivg.au.tsinghua.edu.cn/~jfeng/index.html)†, [Jie Zhou](https://scholar.google.com/citations?user=6a79aPwAAAAJ&hl=en&authuser=1), [Jiwen Lu](http://ivg.au.tsinghua.edu.cn/Jiwen_Lu/)

\* Equal contribution   † Corresponding author

In this work, we propose an efficient multi-level convolution architecture for <b>3D visual grounding</b>. TSP3D achieves superior performance compared to previous approaches in both <b>inference speed and accuracy</b>.

## Main Results
+ We provide the checkpoints for quick reproduction of the results reported in the paper.
  Benchmark | Pipeline | Acc@0.25 | Acc@0.5 | Inference Speed (FPS) | Downloads |
  :----: | :----: | :----: | :----: | :----: | :----: |
  ScanRefer | Single-stage | 56.45 | 46.71 | 12.43 | [model](https://huggingface.co/gwx22/TSP3D/blob/main/ckpt_scanrefer.pth)

  Benchmark | Pipeline | Acc@0.25 | Acc@0.5 | Downloads |
  :----: | :----: | :----: | :----: | :----: |
  Nr3d | Single-stage | 48.7 | 37.0 | [model](https://huggingface.co/gwx22/TSP3D/blob/main/ckpt_nr3d.pth)
  Sr3d | Single-stage | 57.1 | 44.1 | [model](https://huggingface.co/gwx22/TSP3D/blob/main/ckpt_sr3d.pth)

+ Comparison of 3DVG methods on ScanRefer dataset:
<figure>
<p align="left" >
<img src='intro.jpg' width=400 alt="Figure 2"/>
</p>
</figure>