Method

DGIST Multi Task CNN [DGIST MT-CNN]


Submitted on 12 Nov. 2020 03:06 by
Jaehyeong Park (DGIST)

Running time:0.06 s
Environment:GPU @ 1.0 Ghz (Python + C/C++)

Method Description:
The fully convolution CNN(one-stage) model based
2D object detection model.
Applying for multi-task deep learning approach for
different type detection dataset.
Parameters:
Iteration = 20kx8=160k
clip_norm = 0.85
learning_rate = 0.03
back_bone = resnet50
batch_norm_decay = 0.99
Latex Bibtex:

Evaluation in Bird's Eye View


Benchmark MaxF AP PRE REC FPR FNR
UM_ROAD 94.50 % 93.29 % 94.12 % 94.87 % 2.70 % 5.13 %
UMM_ROAD 96.10 % 95.49 % 95.80 % 96.40 % 4.65 % 3.60 %
UU_ROAD 93.48 % 92.64 % 93.99 % 92.98 % 1.94 % 7.02 %
URBAN_ROAD 94.81 % 93.80 % 93.91 % 95.73 % 3.42 % 4.27 %
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