The KITTI Vision Benchmark Suite

Method

Convolutional Neural Network [CN]

Submitted on 27 Sep. 2013 16:23 by
Jannik Fritsch (Honda Research Institute Europe)

Running time:		2 s
Environment:		1 core @ 2.5 Ghz (C/C++)

Method Description:

The method applies a convolutional neural network to label road scene images. It includes a texture descriptor that learns a linear combination of color planes to obtain maximal uniformity in road areas in the test image. The final classification is obtained by combining acquired (offline) and current (online) image information. Note that this algorithm has not been retrained on the KITTI-ROAD dataset but uses the classifier trained on the original dataset (see publication), resulting in non-optimal performance. Finally, the weights of the color planes for each image have been obtained using a quadratic formulation.

Parameters:

CNN-7

Latex Bibtex:

@INCOLLECTION{Alvarez2012ECCV,
author = {Alvarez, Jose M. and Gevers, Theo and LeCun, Yann and Lopez, Antonio M.},
title = {Road Scene Segmentation from a Single Image},
booktitle = {ECCV 2012},
publisher = {Springer Berlin Heidelberg},
year = {2012},
volume = {7578},
series = {Lecture Notes in Computer Science},
pages = {376-389}
}

Evaluation in Bird's Eye View

Benchmark	MaxF	AP	PRE	REC	FPR	FNR
UM_ROAD	73.69 %	76.68 %	69.18 %	78.83 %	16.00 %	21.17 %
UMM_ROAD	86.21 %	84.40 %	82.85 %	89.86 %	20.45 %	10.14 %
UU_ROAD	72.25 %	66.61 %	71.96 %	72.54 %	9.21 %	27.46 %
URBAN_ROAD	79.02 %	78.80 %	76.64 %	81.55 %	13.69 %	18.45 %

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