PointPillars for KITTI

Description

The nutonomy/second.pytorch repository on GitHub offers a specific implementation of PointPillars for object detection within the KITTI dataset. This project serves as a demonstration to reproduce the findings presented in the paper 'PointPillars: Fast Encoders for Object Detection from Point Clouds,' intended for publication at CVPR 2019. It achieves this by making minimal modifications to the existing open-source codebase, SECOND.

This repository is a fork of the original SECOND codebase, tailored for KITTI object detection tasks. It is important to note that this code is not an official nuTonomy codebase, though it is designed to align with the published PointPillars results. The project has been archived by its owner as of October 13, 2021, and is now in a read-only state. For actively maintained code that can also reproduce PointPillars results on the nuScenes dataset, users are directed to the SECOND repository.

Getting started with this project involves cloning the repository, installing necessary Python packages (including PyTorch, Anaconda, and specific libraries like shapely, pybind11, protobuf, scikit-image, numba, pillow, google-sparsehash, fire, and tensorboardX), and setting up CUDA for Numba. The installation process also includes instructions for SparseConvNet, which is expected to be configured correctly by the general SECOND codebase, and potentially Boost geometry. Environment variables for Numba and PYTHONPATH configuration are also detailed.

Dataset preparation is a crucial step, requiring the download of the KITTI dataset and the creation of specific directory structures. Users must then generate KITTI information files, create reduced point clouds, and generate groundtruth database information using provided Python scripts. Finally, configuration files need to be edited to point to the prepared datasets. The repository outlines the training and evaluation procedures, including commands for single-GPU training and resuming from checkpoints, as well as how to save detection results in official KITTI label format.

PointPillars for KITTI Highlights

• Object detection from point clouds

• Implementation of PointPillars algorithm

• KITTI dataset compatibility

• Fork of the SECOND codebase

• Reproduces published PointPillars results

• Codebase supports Python 3.6+

• Requires PyTorch 0.4.1+

• Tested on Ubuntu 16.04/18.04

• Includes dataset preparation scripts

• Provides training and evaluation scripts

• Single GPU training support

• Batch size of 2 for training

• Archived and read-only repository

Getting Started with PointPillars for KITTI

1. 1

   Clone code: git clone https://github.com/nutonomy/second.pytorch.git

2. 2

   Install Python packages: Use Anaconda for core packages, then pip for others.

3. 3

   Setup CUDA for Numba: Configure environment variables in ~/.bashrc.

4. 4

   Add to PYTHONPATH: Include second.pytorch/ in your PYTHONPATH.

5. 5

   Prepare dataset: Download KITTI, create directories, and run data preparation scripts.

6. 6

   Modify config file: Update paths to datasets in configuration files.

7. 7

   Train model: Run the training script with specified config and model directories.

8. 8

   Evaluate model: Execute the evaluation script to assess model performance.

PointPillars for KITTI's Use Cases

• Object Detection
• Autonomous Driving Research
• Point Cloud Analysis
• Computer Vision Benchmarking
• Academic Research

FAQ from PointPillars for KITTI

• What is PointPillars for KITTI object detection?

  PointPillars for KITTI object detection is a GitHub repository that provides code to reproduce the results of the PointPillars algorithm for object detection on the KITTI dataset. It is based on the SECOND codebase.

• Is this repository actively maintained?

  No, this repository was archived by the owner on October 13, 2021, and is now read-only. For actively maintained code, users are recommended to use the SECOND repository.

• What are the system requirements for this code?

  The code supports Python 3.6+ and PyTorch 0.4.1+. It has been tested on Ubuntu 16.04/18.04 and requires CUDA setup for Numba.

• How do I prepare the KITTI dataset for this project?

  You need to download the KITTI dataset, create specific directory structures, and then run provided Python scripts to generate information files, reduced point clouds, and groundtruth database information.

• Can I train a new model with this repository?

  Yes, you can train a new model by running the training script. If the specified model directory does not exist, training will start anew. If it exists, training will resume from the last checkpoint.

• Does this repository support multi-GPU training?

  No, the training process described in the repository only supports a single GPU.

• What is the typical training time for this model?

  On a single 1080Ti GPU, training the xyres_16 configuration for 160 epochs requires approximately 20 hours.

• How are detection results saved?

  By default, results are saved as a result.pkl file. You can save them in the official KITTI label format by using the --pickle_result=False flag during evaluation.

• What is the relationship between this repository and the SECOND codebase?

  This repository is a fork of the SECOND codebase, with modifications specifically for implementing and demonstrating PointPillars for object detection on the KITTI dataset.

• Where can I find the original PointPillars paper?

  The paper is available on arXiv at https://arxiv.org/abs/1812.05784v1.