Mask R-CNN for Segmentation of Aerial Data with Edge Aware Loss
Authors
Hansen, Frederik Østerby ; Holst, Matias Dahlin ; Hem, Mikkel Vestergaard
Term
4. term
Education
Publication year
2020
Pages
8
Abstract
Dette projekt undersøger billedsegmentering i luftdata med fokus på at tegne præcise segmenteringsmasker omkring bygninger. Segmenteringsmasker er pixelnøjagtige omrids, der markerer, hvilke pixels der tilhører en bygning. Vi anvender Mask R-CNN, en udbredt computer vision-model, der kan finde objekter og forudsige en maske for hvert objekt. Vores hovedidé er at forbedre, hvordan modellen lærer at tegne masker, ved at ændre maskens tabsfunktion (loss-funktion) – den regel, der måler fejl under træningen. Vi implementerer to alternative tabsfunktioner og undersøger deres effekt på nøjagtigheden, når de bruges på vores eget datasæt. Datasættet er sammensat af luftfotografisk data og LiDAR-data (laserbaserede 3D-målinger). Resultaterne viser, at den Edge Aware-tabsfunktion giver en bemærkelsesværdig forbedring af kvaliteten af de genererede masker.
This project studies image segmentation in aerial data with the goal of drawing precise segmentation masks around buildings. Segmentation masks are pixel-level outlines that mark which pixels belong to a building. We use Mask R-CNN, a widely used computer vision model that detects objects and predicts a mask for each one. Our main idea is to improve how the model learns to draw masks by changing the mask loss function—the rule that measures errors during training. We implement two alternative loss functions and examine their effect on accuracy when applied to our own dataset. The dataset is derived from aerial photographic data and LiDAR data (laser-based 3D measurements). Our results show that the Edge Aware loss function produces a noteworthy improvement in the quality of the output masks.
[This abstract was generated with the help of AI]
Documents