Method

LightFusion: Lightweight CNN Architecture for Enabling Efficient Sensor Fusion in Autonomous Driving [LightFusion]
[Anonymous Submission]

Submitted on 14 Dec. 2022 14:41 by
[Anonymous Submission]

Running time:0.027 s
Environment:GPU @ 1.5 Ghz (Python)

Method Description:
We propose a framework, named LightFusion, to
develop one lightweight and accurate CNN
architecture (LA-RoadNet) for enabling efficient
fusion. Firstly, our LA-RoadNet is constructed by
maintaining the same number of fusion stages in
the baseline model and further cutting down the
original number of basic blocks in each fusion
stage of this baseline model, thus greatly
reducing the computational overhead. Then, for
gaining high perception accuracy, we introduce a
joint unbalanced loss for guiding our LA-RoadNet
to both mimic the baseline model’s structured
information and learn from original ground-truth
labels together.
Parameters:
w_task = 0.5
Latex Bibtex:

Evaluation in Bird's Eye View


Benchmark MaxF AP PRE REC FPR FNR
UM_ROAD 96.03 % 93.35 % 95.45 % 96.63 % 2.10 % 3.37 %
UMM_ROAD 97.69 % 95.57 % 97.42 % 97.96 % 2.85 % 2.04 %
UU_ROAD 95.04 % 92.61 % 94.22 % 95.88 % 1.92 % 4.12 %
URBAN_ROAD 96.53 % 93.87 % 96.03 % 97.03 % 2.21 % 2.97 %
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