Home

Awesome

PG-RCNN: Semantic Surface Point Generation for 3D Object Detection (ICCV 2023)

<p align="center"> <img src="docs/PGRCNN_video.gif" width="95%"> </p> <p align="center"> <img src="docs/PGRCNN_video_bev.gif" width="95%"> </p>

This is the official implementation of "PG-RCNN: Semantic Surface Point Generation for 3D Object Detection" (ICCV 2023).

<p align="center"> <img src="docs/framework.png" width="95%"> </p>

[Paper] [Supp] [ArXiv]

Thanks to OpenPCDet, our implementation is based of pcdet v0.5.2.

Overview

Model Zoo

The results are the 3D detection performance of moderate difficulty on the val set of KITTI dataset.

training timeCar@R40Pedestrian@R40Cyclist@R40download
PGRCNN~4.5 hours85.2558.3775.04model-8.8M

Note that the performance may vary a little due to sampling in PointNet++ encoder.

Installation

Please refer to INSTALL.md for the installation of OpenPCDet.

To train PG-RCNN, You need to additionally install pytorch3d for utilizing Chamfer Distance. We recommend using pytorch3d ver0.7.0.

Getting Started

Please refer to GETTING_STARTED.md to learn more usage about this project.

Important: Generate approximated complete object points

Under pcdet directory, execute:

python -m pcdet.datasets.multifindbestfit

License

PG-RCNN is released under the Apache 2.0 license.

Acknowledgement

We would like to thank the authors of OpenPCDet and BtcDet for their open source release of their codebase.