Method

Self-supervised PointGNN [SSL-PointGNN]
will be updated after the acceptance of the paper

Submitted on 1 Feb. 2022 10:36 by
Mingyu Liu (Technische Universität München)

Running time:0.56 s
Environment:GPU @ 1.5 Ghz (Python)

Method Description:
We propose a self-supervised (SSL) generic point cloud pre-training















method to improve the 3D object detection task quality. This is the















first study that shows learned motion and flow representations in a















spatio-temporal context via self-supervised training, which can be















utilized for downstream spatial 3D vision tasks. We submit our















PointGNN results trained with our SSL pipeline.
Parameters:
1400k step training, T=3, PointGNN car network
Latex Bibtex:


@article{erccelik20223d,

title={3D Object Detection with a Self-supervised Lidar Scene Flow
Backbone},

author={Erçelik, Emeç and Yurtsever, Ekim and Liu, Mingyu and
Yang, Zhijie and Zhang, Hanzhen and Topçam, Pınar and Listl,
Maximilian and Çaylı, Yılmaz Kaan and Knoll, Alois},

journal={arXiv preprint arXiv:2205.00705},

year={2022}

}



journal={arXiv preprint arXiv:2205.00705},



year={2022}



}

Detailed Results

Object detection and orientation estimation results. Results for object detection are given in terms of average precision (AP) and results for joint object detection and orientation estimation are provided in terms of average orientation similarity (AOS).


Benchmark Easy Moderate Hard
Car (Detection) 96.61 % 93.65 % 88.53 %
Car (Orientation) 38.55 % 37.21 % 36.53 %
Car (3D Detection) 87.78 % 79.36 % 74.15 %
Car (Bird's Eye View) 92.92 % 89.16 % 83.99 %
This table as LaTeX


2D object detection results.
This figure as: png eps pdf txt gnuplot



Orientation estimation results.
This figure as: png eps pdf txt gnuplot



3D object detection results.
This figure as: png eps pdf txt gnuplot



Bird's eye view results.
This figure as: png eps pdf txt gnuplot




eXTReMe Tracker