Developing of OpenPCDet

This repository forked from OpenPCDet, aiming to develop some algorithms and document concise and pratical tutorials, so as to use it simply.

Running Environment

• Ubuntu 20.04

• OpenPCDet v0.6.0

Introduction

Official document can be referred to here.

What does OpenPCDet toolbox do?

OpenPCDet is a general PyTorch-based codebase for 3D object detection from point cloud. It currently supports multiple state-of-the-art 3D object detection methods with highly refactored codes for both one-stage and two-stage 3D detection frameworks.

Based on OpenPCDet toolbox, OpemMMLab win the Waymo Open Dataset challenge in 3D Detection, 3D Tracking, Domain Adaptation three tracks among all LiDAR-only methods.

OpenPCDet design pattern

• Data-Model separation with unified point cloud coordinate for easily extending to custom datasets:

Here are a few thoughts (questions) in response to this pattern.

1. How do they realize the separation of data and model? What are the shortcomings in previous or other works?

   Simply speaking, the separation refers to decompart the codes of building dataset loader and creating model networks. Take concrete codes as an example, i.e. separation_of_data_and_model.ipynb. Fast training and testing of datasets such as kitti, nuscenes, waymo, etc., can be achieved through the interface provided by the pcdet toolkit, and the network parameters can also be modified by modifying the yaml file prior to training, a programming specification which is of great help to us. In the tools directory, demo.py is fine example, by inheriting DatasetTemplate, it is possible to read and infer sample data from different datasets.

2. What is the unified point cloud coordinate, and why are they doing it?

   #Issue 236 raised an question about "How does deal with the ".pcd" point cloud files to the unified normative coordinate of OpenPCDet?". With the help of this question, it is just necessary to understand what the unified point cloud coordinates are in the pcdet toolbox.

3. How does this toolbox extent to custom datasets, such as Carla-Custom dataset or physical vehicle's dataset?

   As for Carla, there are many public projects that is trying to collect data with KITTI format using Carla simulator, such as Mines Paristech's NPM3D Benchmark Suite, it can produce KITTI-Carla dataset, a KITTI-like dataset generated by CARLA Simulator. If we can use this custom dataset to train an network. It may be possible to get more persuasive in the article. What's more, we can also rely on physical vehicles to collect data from real scenarios and produce dataset like CODA.

...