The Figure 4 shows nine levels with respect to data quality in
the realm of automated driving. A full suite of nine levels
might not be appropriate in all scenarios. Level 1 is the
specific setup, can be seen as a configuration parameter for
the dataset. Two data checks, Level 2 and 4 are data checks
pre- & post- the data collection. Errors in these two are noted,
but not used in the final DRL calculation here, future work
will, however. Level 3 is impairments from the test drive
itself: examples are loose sensors, vibration-induced artifacts,
occlusion from other vehicles, equipment failures, sensing
errors, communications and processing to name just a few
detractors. Correlating IMU data with some modalities will
be observable, but not all. A simple measure we use is the
percentage of useful frames (in relation to all). Level 5 is
a physical impairment, that will affect sensors, particularly in
adverse weather, but is not explicitly accounted for in
adverse weather, but is not explicitly accounted for in
this paper. Our datasets, explained in Section V are taken
in harsh or adverse Nordic conditions, deliberately to test
sensor impairments. Levels 6 and 7 are the levels discussed
in this paper. Each modality is handled separately, and each
modality has several tools that are scaled and combined to
form an aggregate score for these layers. We neither study
object recognition nor HD MAPS that are covered in Levels
8 and 9 in this paper, but are slated for future work. Note,
DRL layers depend on lower layers, indicated by the up
arrows under each DRL level. That means, as well as the
important research such as machine learning algorithms for
object detection, issues such as clean and calibrated sensors
will have a profound impact on the quality upstream

Levels 6-7 in the DRL structure hold the main contribution
within this paper. We initially selected 40 (IQA) and 10
point-cloud (PC-IQA) quality tools. Throughout the results
section, we narrow them down. Each metric from the selec-
tion ’scores’ images and point clouds from 2 datasets. The
scores are combined, downsampled and processed into a time
series, and then averaged, the algorithm shown in Figure 9.
In this work, the α, β, γ, δ parameters are assigned equal
weights. λ, or decay scalar is set to 0.9.
That is 90% weight on the past values and 10% on the recent
scores. This value tracks score fluctuations, reasonably well
without over- or undershooting the score tracking. A detailed
investigation could be done or make it time dependent λ(t).
The sampling rate r is set to 10 Hz, the LiDAR scan rate.