Advanced water scene image augmentation method for deep learningKoželj, Janja (Avtor) Kristan, Matej (Mentor) Zavrtanik, Vitjan (Komentor) computer visionimage augmentationdiffusion modelsMore data is required to train better models for various downstream tasks to prevent overfitting and improve generalization, which is often solved by using data augmentation techniques. Because creating large domain specific semantic segmentation datasets is challenging, they are often smaller and contain limited variety of conditions and environments. In this work, we present advanced data augmentation methods, to expand dataset size, quality, and diversity by leveraging large diffusion models capable of introducing novel conditions and environments into the original dataset. We utilize diffusion models to create multiple data augmentation pipelines, apply them on MaSTr1325 dataset, and evaluate performance of a semantic segmentation model trained on the augmented datasets on two other maritime object detection benchmarks. The augmentation methods improve water-edge accuracy and achieve comparable object detection performance on MODS benchmark. Additionally, the detection performance and segmentation accuracy improves when using presented data augmentation methods on a more complex LaRS dataset, most notably the MaskAugment method, which shows a 2.1% mIoU improvement to 93.0% and a 10.5% improvement in F1 score to 60.2% compared to the baseline without advanced augmentation.20242024-11-29 12:25:01Magistrsko delo/naloga165293sl