SyncMapV2 is an unsupervised learning-based image segmentation algorithm that demonstrates remarkable robustness against digital impairments (e.g., noise, weather effects, cloudiness, etc.) compared to existing state-of-the-art (SOTA) algorithms. While existing SOTA algorithms show significant decreases in mean IoU (mIoU) against digital impairments (e.g., 37.7% for noise, 33.8% for weather, and 29.5% for cloudiness), SyncMapV2 achieves only 0.01% mIoU decrease. This performance is achieved through a learning paradigm that combines self-organizing dynamics equations and random network concepts, without robust training, supervised learning, or special loss functions. Furthermore, unlike existing algorithms, it adapts online without the need for reinitialization for every input, mimicking the continuous adaptability of human vision. Adaptability tests also show virtually no performance degradation.
