The KITTI Vision Benchmark Suite

Method

Deep Deconvolutional Networks for Scene Parsing [DDN ]

Submitted on 23 Nov. 2014 01:02 by
Rahul Mohan (Stanford University)

Running time:		2 s
Environment:		GPU @ 2.5 Ghz (Python + C/C++)

Method Description:

This method uses deep deconvolutional neural networks, a new
deep learning architecture that improves on traditional
convolutional neural networks. Deconvolutional layers are
stacked on top of traditional convolutional layers.

In addition, a multi-patch training technique is used, where
multiple networks are trained on various parts of the input
image, in order to learn an effective spatial prior.

Parameters:

num_patches=16

Latex Bibtex:

@misc{Mohan2014ARXIV,
Author = {Rahul Mohan},
Title = {Deep Deconvolutional Networks for Scene Parsing},
Year = {2014},
Eprint = {arXiv:1411.4101},
}

Evaluation in Bird's Eye View

Benchmark	MaxF	AP	PRE	REC	FPR	FNR
UM_ROAD	93.65 %	88.55 %	94.28 %	93.03 %	2.57 %	6.97 %
UMM_ROAD	94.17 %	92.70 %	96.73 %	91.74 %	3.41 %	8.26 %
UU_ROAD	91.76 %	86.84 %	93.06 %	90.50 %	2.20 %	9.50 %
URBAN_ROAD	93.43 %	89.67 %	95.09 %	91.82 %	2.61 %	8.18 %

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.

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