The KITTI Vision Benchmark Suite

Method

LoDNN [la] [LoDNN]

Submitted on 8 Jan. 2017 01:26 by
Luca Caltagirone (Chalmers University of Technology)

Running time:		18 ms
Environment:		GPU @ 2.5 Ghz (Torch)

Method Description:

FCN for road segmentation in LIDAR top-view images.

Parameters:

lr = 0.001
drop = 0.25

Latex Bibtex:

@article{CaltagironeEtAl2016,
title={Fast LIDAR-based Road Detection
Using Fully Convolutional Neural Networks},
author={Caltagirone, Luca and
Scheidegger,
Samuel and Svensson, Lennart and Wahde, Mattias},
journal={IEEE Intelligent Vehicles
Symposium},
year={2017}
}

Evaluation in Bird's Eye View

Benchmark	MaxF	AP	PRE	REC	FPR	FNR
UM_ROAD	92.75 %	89.98 %	90.09 %	95.58 %	4.79 %	4.42 %
UMM_ROAD	96.05 %	95.03 %	95.79 %	96.31 %	4.66 %	3.69 %
UU_ROAD	92.29 %	90.35 %	90.81 %	93.81 %	3.09 %	6.19 %
URBAN_ROAD	94.07 %	92.03 %	92.81 %	95.37 %	4.07 %	4.63 %

This table as LaTeX

Behavior Evaluation

Benchmark	PRE-20	F1-20	HR-20	PRE-30	F1-30	HR-30	PRE-40	F1-40	HR-40

This table as LaTeX

Road/Lane Detection

The following plots show precision/recall curves for the bird's eye view evaluation.

This figure as: png eps pdf

Distance-dependent Behavior Evaluation

The following plots show the F1 score/Precision/Hitrate with respect to the longitudinal distance which has been used for evaluation.

Visualization of Results

The following images illustrate the performance of the method qualitatively on a couple of test images. We first show results in the perspective image, followed by evaluation in bird's eye view. Here, red denotes false negatives, blue areas correspond to false positives and green represents true positives.

This figure as: png