Method

3D car detection based on refinedet [ref3D]


Submitted on 18 Apr. 2019 07:08 by
Yichuan Miao (Beijing University of post and telecommunacation)

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

Method Description:
3D object detection and pose estimation from a
single image or an video, using a deep
convolutional neural network to estimate the
dimention and local ray. Using the refinedet
trained by kitti dataset to detect the 2d
bounding box of car. Then combine with the 2D
bounding box to choose the best location of the
car which has the lowest error score.
Parameters:
batchsize = 16
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) 74.81 % 57.14 % 47.95 %
Car (Orientation) 74.41 % 56.54 % 47.15 %
Car (3D Detection) 0.00 % 0.00 % 0.00 %
Car (Bird's Eye View) 0.00 % 0.00 % 0.01 %
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