Method

Auxiliary multi scale deep learning network for road detection [DGIST-AMSL]


Submitted on 18 Jun. 2020 03:12 by
Woong-Jae Won (Daegu Gyeongbuk Institute of Science & Technology)

Running time:0.04 s
Environment:GPU @ 2.5 Ghz (Python)

Method Description:
Apply for auxiliary multi-scale training method for
decovolution layer road segmentation. And, we also
apply for multi-scale training rather then geometry
augmentation
Parameters:
learning_rate 0.015
iteration 80000
multi-scale(input): 13 level
Latex Bibtex:

Evaluation in Bird's Eye View


Benchmark MaxF AP PRE REC FPR FNR
UM_ROAD 94.12 % 88.74 % 94.49 % 93.76 % 2.49 % 6.24 %
UMM_ROAD 96.35 % 93.02 % 97.08 % 95.63 % 3.16 % 4.37 %
UU_ROAD 94.45 % 89.28 % 95.75 % 93.19 % 1.35 % 6.81 %
URBAN_ROAD 95.23 % 90.49 % 95.98 % 94.48 % 2.18 % 5.51 %
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

This figure as: png eps pdf

This figure as: png eps pdf

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

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png


eXTReMe Tracker