Method

Graph-Voxel Feature Fusion Network [GV-RCNN]
www.njust.edu.cn

Submitted on 1 Jun. 2022 14:47 by
Huang Tian Tian (Nanjing University of Science and Technology)

Running time:0.1 s
Environment:1 core @ 2.5 Ghz (Python + C/C++)

Method Description:
we use a graph network to solve the quantization
loss problem and provide rich local detail
information to the model and use a point cloud
segmentation network to supervise the graph
network learning. In addition, a multi-level
feature fusion method is applied in the model,
which effectively improves the representation of
the feature of the 3D proposals.
Parameters:
\alpha=0.2
Latex Bibtex:

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) 96.29 % 95.06 % 92.43 %
Car (Orientation) 96.28 % 94.96 % 92.28 %
Car (3D Detection) 90.31 % 81.75 % 77.17 %
Car (Bird's Eye View) 94.52 % 88.94 % 86.24 %
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