Pothole Detection Using Deep Learning: A Real-Time and AI-on-the-Edge Perspective

Table 4

Comparison with state-of-the-art vs. our results.

Contribution	Dataset	[email protected]	[email protected]	Inference time (ms)
Omar et al. [43]	Pothole image dataset (PID)	60%	—	—
Shaghouri et al. [44]	Pothole image dataset (PID)	75.53%	—	—
Gajjar et al. [45]	Self + Online collected	—	18.5%	481
Sung-Sik et al. [46]	Pothole image dataset (PID)	—	74.8%	—
Our trained SSD-Mobilnetv2	Pothole image dataset (PID)	—	47.4%	7
Our trained YOLOv3	Pothole image dataset (PID)	83.60%	—	70.57
Our trained YOLOv4	Pothole image dataset (PID)	85.48%	—	52.51
Our trained YOLOv5	Pothole image dataset (PID)	—	95%	10