Skip to content

Instantly share code, notes, and snippets.

@ddksr
Created February 7, 2015 16:11
Show Gist options
  • Save ddksr/be0d6fa187bc2d004d5f to your computer and use it in GitHub Desktop.
Save ddksr/be0d6fa187bc2d004d5f to your computer and use it in GitHub Desktop.
CVWW presentation

Highlights

  • multi-image photogrammetry is becoming the most accurate and cost effective method of documentation in underwater archeology
  • a 3d point cloud is difficult to use directly for analysis and comparison of object shapes
  • volumetric shape models can address some of this difficulties
  • superquadric volumetric shape models can be automatically recovered from a point cloud
  • superquadric models offer an abstraction level suitable for reasoning about a scene

Intro

  • stereo and photogrammetry in 1960s George Bass
  • 3D available only after the field work was already finished
  • documentation performed mostly manually - imperfections
  • 3D points with textures excellent for visualization, not for further analysis

Documentation

  • rigid grid of squares 2m x 2m, 34 diving hours
  • 900 photos, 1h
  • DSRL Nikon 300

Site description

  • Roman shipwreck with sarcophagi cargo 2nd century
  • Sutivan, Brač, Croatia
  • 30 tons
  • largest in west cost Adriatic (besides another a bit norhter)
  • first discovered by PIK Mornar, 2008
  • research 2010-2012
  • photo Rok Kovačič, 3D Gregor berginc Xlab, 3dimenzija, PHOV
  • 40m2, 35m depth
  • 24 semi-manifactured stone blocks, 2 sarcophagi with lids, circular stone column, oil jar
  • ship had to be 20-22 meters
  • antique, wood artefacts dated to 2nd century
  • Aegan basin and Asia Minor

Modeling and segmentation

  • a1, a2, a3 - distances on axies
  • e1, e2 - shape
  • 6 additional for translation and rotation
  • recovery and segmentation at the same time, procedure
  • 3d points covered with small sq seeds
  • each seed fitted to corresponding 3d set
  • the sq models can the next iteration expand so that 3D points that are in the vicinity of the corresponding model and are compatible with the shape of the superquadric can be integrated into the model
  • when superquadric models expand as allowed by the 3D points, they start to overlap
  • a selection procedure is performed using the criterion of minimum description length so that in each iteration fewer superquadric models remain
  • after a few iterations only as many superquadric models remain as are necessary by the parts structure in the 3D data

Results

  • recovered 3, 5, 6, 8, 9, 10, 11 and 13
  • average error, error std deviation
  • compared to manual point-to-point measurements

Conclusions

  • multi-image photogrammetry will prevail
  • sq offer a level of abstraction, important when you have to deal with that much data
  • simple parametrization: very few parameters
  • could be used for shape indexing and and searching for similar objects
  • sq recover average shape of model
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment