ERCIM News No.44 - January 2001 [contents]
ERCIM News No.44 - January 2001 [contents]
by Michal Haindl and Pavel SlavĚk
The purpose of this Virtual Reality project was to build a virtual gallery model from real image data measurements utilising results from the VIRTUOUS project. The resulted virtual model of the Department of Modern Art of the National Gallery in Prague is intended to serve as the core of the web based National Gallery information system.
Customary manual creation of virtual reality models of real world scenes is tedious and error-prone work as the scene complexity increases and any automation can substantially reduce the laboriousness and consequently the cost of the whole process. Visual sensors offer the ideal source of data for such a task especially when range and vision sensors are already common and their mutual registration can be done using either standard photogrammetric techniques or an appropriate sensor setup.
To make virtual worlds realistic detailed scene models must be built. Satisfactory models require not only complex 3D shapes accorded with the captured scene, but also lifelike colour and texture. Textures provide useful cues to a subject navigating in such a VR environment, and they also aid in the accurate detailed reconstruction of the environment. This increases significantly the realism of the synthetic generated scene.
Our pilot system for the virtual model acquisition currently restricted to planar-faced scene objects was used to build the gallery model. The core part of this VR modelling system is a fast range image segmentation algorithm for planar-faced scene objects. The Department of Modern Art, housing a collection of images, drawings and statues from the period of 20th century, is located in a functionalistic building in Prague, where the planar-faced object restriction is quite acceptable. The segmentor combines a discontinuity detection step and the subsequent line based region growing guided by these detected discontinuities. Range segmentation results subsequently supervise the following - geometry inference and texture analysis steps.
The gallery model is composed from mutually connected partial submodels (halls, corridors, etc.) containing mostly planar-faced objects. However although even if a single planar face is correctly found, its corresponding range pixels values are too noisy to be of direct use for the face model construction. Such a model would have many small noise-generated faces which have to be eliminated. Noise filtering is done through the least-square fitting of a plane to a detected face perimeter pixels and only the rectified face pixels are subsequently used for a virtual model face geometry inference.
Thematic maps from the range segmentor serve also as the region maps for the registered colour textured images. If a detected face has sufficient area for a reliable texture model inference then this model is inferred from the corresponding projected colour textured image area, otherwise the face is classified as too small for individual texturing and the texture model can be assigned interactively if required. Textures are modelled using our fast multi-grid Markov random field based model and combined with model geometry information.
Resulting gallery model is quite complex hence it has to be completed with a virtual environment navigation system. Otherwise its user can be easily lost in the labyrinth of gallery virtual corridors. Our solution to this problem is an automatic path generation that defines a trajectory of the virtual walk through. The path parameters (entry and exit points, obstacle avoidance - eg, usage of a staircase for handicapped visitors, etc.) are defined by the user. The process of the path finding is in principle finding a path in a underlying graph describing the virtual scene and constrained by requested art works to be seen on the way.
The virtual model of the Department of Modern Art of the National Gallery in Prague was semi-automatically created in the VRML2 language so it can be easily placed on the web server with sufficient resources. The model recently describes interior public exhibition areas of the gallery building together with the exhibition panels holding paintings from the gallery art collection and can be used not only as the visitors virtual information system but also for gallery workers to model optimal show arrangement for exhibitions to be prepared. Although recent technology advances already enables automatic or at least semi-automatic construction of distributed virtual models, further research is still needed especially in the area of general surface object modelling. Similarly the VRML2 language has many functional restrictions and some better distributed virtual reality modelling language is clearly required.
Links:
VIRTUOUS project: http://www.ee.surrey.ac.uk/EE/VSSP/3DVision/virtuous/virtuous.html
National Gallery demo: http://www.utia.cas.cz/user_data/haindl/PR/demHP.html
Please contact:
Michal Haindl - CRCIM (UTIA)
Tel: +420 2 6605 2350
E-mail: haindl@utia.cas.cz