PointSlice

This repository contains the official implementation of the PointSlice model, offering an innovative 3D object detection method. You can use the following approaches to utilize it.

Installation Guide

This document provides a step-by-step guide to installing the required dependencies and setting up the environment for running the project.

Requirements

The code has been tested in the following environments:

• Operating System: Linux (Ubuntu 20.04)
• Python: 3.6+
• PyTorch: 1.1 or higher
• CUDA: 9.0 or higher (PyTorch 1.3+ requires CUDA 9.2+)
• Sparse Convolution Library: spconv
  • spconv v1.0 (commit 8da6f96) for PyTorch 1.1
  • spconv v1.2 for PyTorch 1.3+
  • spconv v2.x (latest version, install via pip)

For detailed package versions, please see our requirements.txt.

Install pcdet

Please install the pcdet library and its dependencies by running:

python setup.py develop

Usage

To begin, please follow the HEDNet Getting Started Guide to download and prepare the dataset. Note that this process requires significant disk space and time.

Once the data is ready, you can run the code using the following command:

Waymo

bash scripts/dist_train.sh cfgs/pointslice/pointslice_1f_1x_waymo.yaml 8 --batch_size 16 --epoch 24 --workers=2

You can evaluate the model's performance using the following command:

python test.py --cfg_file cfgs/pointslice/pointslice_1f_1x_waymo.yaml --ckpt {yourckpt.pth path} --batch_size 1

We provide a pretrained model trained on the Waymo dataset. You can download the checkpoint (checkpoint_epoch_24.pth) from the link below:

Waymo Download Link: Waymo | Hugging Face

Argoverse2

You can download the checkpoint from the link below:

Download Link: Argoverse2 | Hugging Face

nuScenes

For nuScenes code, you can refer to this GitHub repository.

Acknowledgement

Our code is based on OpenPCDet and HEDNet. We thank the authors for their open-source contribution.