The KITTI Vision Benchmark Suite

Method

Hierarchal Inference Machine [HIM]

Submitted on 21 May. 2015 00:46 by
Ankit Laddha (Carnegie Mellon University)

Running time:		7 s
Environment:		>8 cores @ 2.5 Ghz (Python + C/C++)

Method Description:

A hierarchical Graphical Model for an image is build. For each layer
in the hierarchy a Random forest is trained to classify.

Parameters:

Number of trees = 10
Max Depth of Trees = 10
layers in hierarchy = 7

Latex Bibtex:

@inproceedings{Munoz2010ECCV,
author = "Daniel Munoz and J. Andrew Bagnell and Martial
Hebert",
title = "Stacked Hierarchical Labeling",
booktitle = "European Conference on Computer Vision (ECCV)",
year = "2010"
}

Evaluation in Bird's Eye View

Benchmark	MaxF	AP	PRE	REC	FPR	FNR
UM_ROAD	90.07 %	79.98 %	90.79 %	89.35 %	4.13 %	10.65 %
UMM_ROAD	93.55 %	90.38 %	94.18 %	92.92 %	6.31 %	7.08 %
UU_ROAD	85.76 %	76.18 %	87.65 %	83.95 %	3.86 %	16.05 %
URBAN_ROAD	90.64 %	81.42 %	91.62 %	89.68 %	4.52 %	10.32 %

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