Table 1.
Cadamuro et al. results.
Fig 1.
Plot of the average vibrations, , for an individual device in the order of time collected, corresponding to dates between September 4th and November 26th, 2020. In this case,
in the direction of the center-of-Earth to sky axis was used. Each point represents a road segment. The quality thresholds for this setup were determined to be 0.375 and 0.600, where points falling below
represent high quality roads, those above
are low quality, and points in-between are mid-quality. The spike on the right corresponds to a particular drive on dirt roads in the Appalachians near Blacksburg, VA, USA.
Fig 2.
Random sample of cropped road segment images resized to 200x200 with corresponding labels. Values in the range [0–1] correspond to low quality roads, [1–2] corresponds to mid-quality, and [2–3] for high quality.
Fig 3.
Example of the determination of the continuous quality value (CQV) of a road segment for a particular device setup. The line is split into 3 parts, representing each of the three quality classes. The slope for the [2–3] quality range is based on the CQV = 2.0 threshold value for the device setup and the constraint of the upper range, CQV = 3.0. The slope for the [1–2] range is determined from the CQV = 1.0 and CQV = 2.0 thresholds. Finally, the slope of the [0–1] range is the average of the combined range [1–3]. Negative CQVs (i.e., roughness values greater than the maximum defined) were re-specified to 0. In this example, the threshold between high and mid-quality roads, , is 0.125, and the threshold between mid and low,
, is 0.28. The colors red, orange, and green correspond to the regions of low, mid, and high quality roads, respectively.
Fig 4.
Example of a CQV of 1.60 indicating a road segment in the mid-quality class. The probabilities for the label were based on the proportion of the area under the curve of a normal distribution (σ = 0.25) that falls into each of the three classes. In the case of this road segment, the label is [0.008,0.937,0.055].
Table 2.
Architecture results.
Table 3.
Ensemble results.
Fig 5.
On the left, a 3x3 group of correct predictions with associated labels and, on the right, a 3x3 group of incorrect predictions with predictions and labels.
Fig 6.
The above shows two random correctly classified test images in the first column. The second, third, and forth columns show the pixels/features that contributed against and in favor of prediction for each of the three classes. For example, the upper-left image is a high quality-labeled road that was predicted by the model most likely to be a high quality road, then mid, then low. Blue pixels represent areas that work against classification in a given class and the red pixels represent areas that work for classification. In this case, the VGG16-based network was investigated.
Table 4.
Nigeria results.