Video recording of the “Auto-labeling of point cloud sequences for 3D object detection” presentation at the “Building reliable datasets for autonomous vehicles” workshop of ITSC 2023.

In this workshop, we introduced the key concepts of our auto-labeling framework, MS3D++. We also give a brief tutorial about how you can use MS3D++ to label your own point clouds.

The content for this workshop is contained in our paper (https://arxiv.org/abs/2308.05988).

GitHub: https://github.com/darrenjkt/MS3D
Tutorial jupyter notebook: https://github.com/darrenjkt/MS3D/blob/main/tools/demo/ms3d_demo_tutorial.ipynb
TLDR explanation video: https://youtu.be/7o5F4AMaIWo?si=az0U_zuQJY1gnz3j

Sections:
0:00 Introduction
3:07 Problems with off-the-shelf 3D detectors
7:40 MS3D++ auto-labelling framework overview
8:31 Multi-source detector ensembling
12:13 Temporal refinement
15:31 Iterative Label Refinement
16:50 Results
19:04 Visualizations
20:21 Tutorial on how to auto-label your own point clouds
24:55 Demo of ouster 128-beam trained with MS3D++ pseudo-labels
25:36 Q & A

source