Andreas Geiger (MPI Tübingen) | Philip Lenz (KIT) | Christoph Stiller (KIT) | Raquel Urtasun (University of Toronto)

Method

Single-stage 3D target vehicle detection based on attention mechanism [AM-SSD]


Submitted on 24 Jan. 2021 16:55 by
shengchao dong (长沙理工大学)

Running time:0.04 s
Environment:1 core @ 2.5 Ghz (Python)

Method Description:
This method adopts the mainstream single-stage
detection, and creatively applies the attention
mechanism network to the vehicle target detection
of 3D point cloud under the premise of not losing
accuracy and speed, in order to expect a
wonderful effect.
Parameters:
lr=0.03 weight_decay=0.01
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) 96.78 % 93.58 % 90.61 %
Car (Orientation) 96.56 % 93.18 % 90.13 %
Car (3D Detection) 89.58 % 80.30 % 75.02 %
Car (Bird's Eye View) 95.56 % 89.74 % 84.65 %
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