Wednesday, July 7, 2010

2nd Day of ISUVR 2010 - Keynote

The 2nd day of ISUVR started with an invited talk by Anton van den Hengel (The Australian Centre for Visual Technologies) The title was Image-based modelling for augmenting reality.

He said that the current mobile AR browsers shows information overlay on the real world, but the information is not registered to the geometry.
Anton claims that User Created Contents (UCC) are required for ubiquitous AR. For AR 3D contents are better, but there haven't been good UCC tools. With current tools epic effort is required to create 3D contents.

Image-based modeling is a good way to create 3D contents because images contains many queues for 3D modeling. There are two way for image-based 3D modeling. Automatic methods generates 3D models of everything, like a 3D laser scanner. Interactive methods allows user to choose the object to interact with.

He introduced Video Trace, which is an interactive image-based 3D modeling system. It is a very nice interactive tool for 3D modeling (However, from my viewpoint, it requires a little bit labor, moving among video frames, modifying 3D geometries..., to finish modeling).

To help 3D modeling, several features and techniques can be exploited:
  • Lines and curves for geometry modeling
  • Mirroring, extrusion, and dense meshing makes modeling an object easier.
  • Surface fitting helps to align planar surfaces that are on the same infinity plane.

Video trace works on a recorded video seuqence. The same idea is extended to the Live AR modeling, so called In-situ modeling. The video below explains the concept and implementation well.

Anton also discussed that there exists misalignment between the modeled real objects and the synthesized virtual ones. It causes visual defect on the occlusion boundary. This problem can be addressed by graph-cut based segmentation method exploiting color distributions.

Video trace requires much user interaction and sometimes they are painful when working with complex 3D objects. To reduce user interaction, silhouette modeling method is adopted. Silhouette based modeling followed by segmentatation generate a 3D model from the video without interaction.

Very nice talk, thank you Anton !

No comments:

Post a Comment