CREATING AND USING MASK IMAGES FOR SEGMENTATION IN POINT CLOUD DATA


Creative Commons License

Ozturk B., Özkar Kabakçıoğlu M.

24th ISPRS Congress on Imaging Today, Foreseeing Tomorrow, Nice, France, 6 - 11 June 2022, vol.43-B2, pp.1133-1138 identifier identifier

  • Publication Type: Conference Paper / Full Text
  • Volume: 43-B2
  • Doi Number: 10.5194/isprs-archives-xliii-b2-2022-1133-2022
  • City: Nice
  • Country: France
  • Page Numbers: pp.1133-1138
  • Keywords: Architectural Heritage, Real-World Data Analysis, Photogrammetry
  • Istanbul Technical University Affiliated: Yes

Abstract

The task of preparing the training data for machine learning is tedious but crucial for accurate results. Aiming at labelling meaningful features semi-automatically rather than manually in order to reduce time, we hereby present initial results for two cases of 13th century Seljuk brick-ornamentation. As our broader research involves machine learning methods for the segmentation of digital survey data for creating meaningful three-dimensional models, the primary goal here is to determine the parts of the patterns from the whole composition and to use this data for different buildings of the genre. Prior to any machine learning, labelling the data of either a whole pattern or pieces of a pattern is a time-consuming task prone to errors. We propose a semi-automated mask generating model for labelling. In order to create the black and white mask images of the original photographs, we utilise the colour difference between the pattern parts. Examined samples have at least three visually distinguishable colours that are turquoise, black and natural. We use photogrammetry-based survey data and image processing to create attributed point clouds and eventually 3D digital models. Using the ready batch processing of a commercial software, we create a distinct mask and apply it to all images of the photogrammetry process. Point cloud data is then created with RAW images, and the generated masks are used to filter desired patterns. As such, we are able to easily label the bricks in the point cloud towards a machine learning training set.