Pixel-accurate road crack detection in presence of inaccurate annotations

Recent road crack detection methods obtain appealing scores but typically allow a few pixel tolerance margin. This is acceptable for locating cracks, not measuring their width (indicator of the cracks’ severity). Our baseline model, U-VGG19, obtains an F-score 71.77% on CrackForest, which superior to other approaches when no admitted. However, increasing without difficult due inaccurate annotations. We propose novel synthetic dataset, Syncrack, as benchmark evaluation training with results show that annotations have detrimental impact F-measure, decreasing it by up 20%. To overcome this, we study label noise correction techniques using weakly supervised learning. Training U-VGG19 these corrected labels improves Syncrack 12%. Obtained CrackForest and Aigle-RN datasets support are useful real-life data too.