PointPillars OpenVINO™ Demo on Intel GPU

Introduction

This repository provides five key demonstrations:

• The implementation is adapted from the Demo of PointPillars Optimization, showcasing how to implement and optimize PointPillars on Intel platforms using OpenVINO™. The original code sources are OpenPCDet, which sets up the PointPillars pipeline demo, and SmallMunich, which converts the PointPillars PyTorch model to ONNX format. For more technical details, refer to the Optimization of PointPillars by using Intel® Distribution of OpenVINO™ Toolkit.

• The repository supports Intel MTL iGPU and Arc770 dGPU platforms.

• It adds an INT8 quantization method for RPN and PFE models.

• Scatter latency statistics are included in the demo outputs.

• Validation has been conducted on Intel MTL iGPU and Arc770 dGPU devices.

---

Overview

This document provides detailed instructions for setting up and running the PointPillars OpenVINO™ Demo on Intel GPU. The sections below include hardware and software requirements, demo setup steps, and quantization guidance.

Sections:

1. Requirements
   • Hardware
   • Software
2. Demo Setup
   • Install iGPU or dGPU Drivers
   • Install Development and Compilation Tools
   • Clone the Repository and Create an Environment
   • Install Python Packages
   • Prepare Datasets
   • Set Up OpenPCDet Package
   • Run the Demo
3. Quantization
   • Prerequisites
   • Update Dataset Path
   • Quantize the Model

---

Requirements

Hardware

Choose one of the following hardware setups:

• Intel MTL with iGPU
• Intel Arc770 dGPU + Core CPU

Software

• Ubuntu 22.04
• Linux Kernel 6.5.0-18-generic
• Python 3.10
• OpenVINO™ 2024.3

---

Demo Setup

1. Install iGPU or dGPU Drivers

1.1 MTL iGPU Driver Installation

Refer to the compute-runtime releases.

1.2 Arc770 Driver Installation

Refer to the Intel Arc GPU documentation.

---

2. Install Development and Compilation Tools

sudo apt update
sudo apt-get install python3-dev
sudo apt install build-essential

---

3. Clone the Repository and Create an Environment

cd /home/shawn
mkdir project
git clone https://github.com/shawn9977/PointPillars-Demo.git
cd PointPillars-Demo
python3 -m venv env_PointPillars
source env_PointPillars/bin/activate

---

4. Install Python Packages

pip install openvino-dev nncf
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

cd /home/shawn/project/PointPillars-Demo
pip install -r requirements.txt

---

5. Prepare Datasets

Refer to SmallMunich's repo for dataset generation.

Note: This step requires an NVIDIA GPU environment. Alternatively, you can skip this step as pre-generated datasets are included in this repository under PointPillars-Demo/main/datasets/training/velodyne_reduced/.

Set the dataset path environment variable:

export my_dataset_path=<your_dataset_folder>/training/velodyne_reduced

# Example:
export my_dataset_path=/home/shawn/project/PointPillars-Demo/datasets/training/velodyne_reduced

---

6. Set Up OpenPCDet Package

python setup.py develop

---

7. Run the Demo

cd /home/shawn/project/PointPillars-Demo/tools/
python demo.py --cfg_file pointpillar.yaml --num -1 --data_path $my_dataset_path

# If `my_dataset_path` is not set, run:
python demo.py --cfg_file pointpillar.yaml --num 100 --data_path /home/shawn/project/PointPillars-Demo/datasets/training/velodyne_reduced

Demo outputs include performance metrics:

INFO  -----------------Quick Demo of OpenPCDet-------------------------
INFO  Loading the dataset and model.
INFO  Number of samples in dataset: xxx
INFO  ------Run number of samples: xxx in mode: balance
INFO  Total: xxx seconds
INFO  FPS: xxx
INFO  Latency: xxx milliseconds
INFO  Scatter latency: xxx milliseconds

INFO  Demo done.

---

Quantization

1. Prerequisites

Complete steps 1-6 from the demo setup.

---

2. Update Dataset Path

Edit quant.py to set the dataset path:

DATA_PATH = "/home/shawn/project/PointPillars-Demo/datasets/training/velodyne_reduced"

---

3. Quantize the Model

Run quant.py to save the quantized model (quantized_pfe.xml) in the tools directory:

cd /home/shawn/project/PointPillars-Demo/tools/
python quant.py