Browsing by Author "Pavoni, Gaia"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Another Brick in the Wall: Improving the Assisted Semantic Segmentation of Masonry Walls
    (The Eurographics Association, 2020) Pavoni, Gaia; Giuliani, Francesca; Falco, Anna De; Corsini, Massimiliano; Ponchio, Federico; Callieri, Marco; Cignoni, Paolo; Spagnuolo, Michela and Melero, Francisco Javier
    In Architectural Heritage, the masonry's interpretation is an essential instrument for analyzing the construction phases, the assessment of structural properties, and the monitoring of its state of conservation. This work is generally carried out by specialists that, based on visual observation and their knowledge, manually annotate ortho-images of the masonry generated by photogrammetric surveys. This results in vectorial thematic maps segmented according to their construction technique (isolating areas of homogeneous materials/structure/texture) or state of conservation, including degradation areas and damaged parts. This time-consuming manual work, often done with tools that have not been designed for this purpose, represents a bottleneck in the documentation and management workflow and is a severely limiting factor in monitoring large-scale monuments (e.g.city walls). This paper explores the potential of AI-based solutions to improve the efficiency of masonry annotation in Architectural Heritage. This experimentation aims at providing interactive tools that support and empower the current workflow, benefiting from specialists' expertise.
  • Thumbnail Image
    Item
    A Validation Tool For Improving Semantic Segmentation of Complex Natural Structures
    (The Eurographics Association, 2019) Pavoni, Gaia; Corsini, Massimiliano; Palma, Marco; Scopigno, Roberto; Cignoni, Paolo and Miguel, Eder
    The automatic recognition of natural structures is a challenging task in the supervised learning field. Complex morphologies are difficult to detect both from the networks, that may suffer from generalization issues, and from human operators, affecting the consistency of training datasets. The task of manual annotating biological structures is not comparable to a generic task of detecting an object (a car, a cat, or a flower) within an image. Biological structures are more similar to textures, and specimen borders exhibit intricate shapes. In this specific context, manual labelling is very sensitive to human error. The interactive validation of the predictions is a valuable resource to improve the network performance and address the inaccuracy caused by the lack of annotation consistency of human operators reported in literature. The proposed tool, inspired by the Yes/No Answer paradigm, integrates the semantic segmentation results coming from a CNN with the previous human labeling, allowing a more accurate annotation of thousands of instances in a short time. At the end of the validation, it is possible to obtain corrected statistics or export the integrated dataset and re-train the network.