Method

3D Single Shot Multiscale FCNN [3D-SSMFCNN]
https://github.com/libornovax/master_thesis_code

Submitted on 7 Jan. 2018 12:54 by
Libor Novak (CTU in Prague)

Running time:0.1 s
Environment:GPU @ 1.5 Ghz (C/C++)

Method Description:
Master's thesis, which describes a deep leanining approach to 2D and 3D bounding box detectioin of cars from monocular images with an end-to-end neural network. The network was created by combining the ideas from DenseBox, SSD, and MS-CNN. It can perform multi-scale detection of 2D or 3D bounding boxes in a single pass and runs in 10fps on 0.5MPx images on a GeForce GTX Titan X GPU.

This is a version with a new 3D bounding box representation.
Parameters:
-
Latex Bibtex:
@mastersthesis{novakmaster2017,
author = {Novak, Libor},
title = {Vehicle Detection and Pose Estimation for Autonomous
Driving},
school = {Czech Technical University in Prague},
year = 2017,
}

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) 77.92 % 78.19 % 69.19 %
Car (Orientation) 77.84 % 77.82 % 68.67 %
Car (3D Detection) 1.88 % 1.41 % 1.11 %
Car (Bird's Eye View) 3.20 % 2.63 % 2.40 %
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