[ICCV2025] Accelerate 3D Object Detection Models via Zero-Shot Attention Key Pruning

This is the implentation for our paper "Accelerate 3D Object Detection Models via Zero-Shot Attention Key Pruning"

中文版本的 README

Introduction

We propose a zero-shot, retraining-free pruning method that accelerates 3D object detection models via key pruning.

News

• [2025/06/26]🔥Our paper is accepted by ICCV 2025 🎉🎉🎉.
• [2025/03/11]🔥We release our code and checkpoints.

Get Started

You need pytorch to use tgGBC.

# 1. install your pytorch.
pip3 install torch torchvision torchaudio

# 2. clone this repo.
git clone https://github.com/iseri27/tg_gbc
cd tg_gbc

# 3. install tggbc
python setup.py develop

Examples

Evaluation

We provide an example of OPEN in examples/open to help you evaluate models with tgGBC.

Training

We provide an example of StreamPETR in examples/streampetr to help you train models with tgGBC. We also provide some checkpoints of StreamPETR-vov-1600x640.

Citation

@InProceedings{xu2025gbc,
    author    = {Xu, Lizhen and Bai, Xiuxiu and Jia, Xiaojun and Fang, Jianwu and Pang, Shanmin},
    title     = {Accelerate 3D Object Detection Models via Zero-Shot Attention Key Pruning},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {23085-23094}
}