: Training these models requires significant GPU memory because of the lack of pooling layers.
: Unlike standard square filters, strip pooling is designed for the elongated, narrow nature of cracks, which helps the model ignore background noise like road markings. Dynamic Loss Functions cracknet github