[vc_row][vc_column width=”1/4″][vc_single_image image=”8041″ css=”.vc_custom_1464177605468{margin-top: 10px !important;}”][/vc_column][vc_column width=”1/2″][vc_column_text]Contact person: Christiano Couto Gava
Funding by: BMBF[/vc_column_text][/vc_column][vc_column width=”1/4″][/vc_column][/vc_row][vc_row][vc_column][vc_column_text]
Reconstruction of 3D-scenes out of camera images represents an essential technology for many applications, such as 3D-digital-cities, digital cultural heritages, games, tele-cooperation, tactical training or forensic. The objective of the project CAPTURE is to develop a novel approach for 3D scene acquisition and develop corresponding theory and practical methods.
[/vc_column_text][/vc_column][/vc_row][vc_row][vc_column][vc_column_text]
3D-scene reconstruction with high resolution and high dynamic range spherical images
Instead of processing a large amount of standard perspective low resolution video images, we use as input data a few single but full spherical high resolution and high dynamic range (HDR) images. Currently available spherical high resolution cameras are able to record fine texture details and the complete scene from a single point in space. Additionally such cameras provide HDR images yielding consistent color and photometric information. We propose to exploit this new technology focusing on the dense/high-quality 3D reconstruction of both indoor and outdoor environments.
The first three images are images of the same object with varying exposure times. The last one is the tonemapped HDR image.
The fundamental issue of the project is to develop novel algorithms that take into account the properties of these images, and thus to push forward the current state of the art in 3D scene acquisition and viewing. In particular we develop novel stable and light-invariant image feature detectors, as well as robust assignment methods for image matching and novel 3D reconstruction/viewing algorithms, which exploit the properties of the images.The multiple spherical view geometry provides a high amount of redundant information about the underlying environment. This, combined with the consistency of the color and photometric information from HDR images, allows us to develop new methods for robust high-precision image matching and 3D structure estimation, resulting in a high-fidelity textured model of the real scene.
The upper three images are three examples of the image series taken for the reconstruction of the dragon. The lower images are results of the reconstruction, whereas the left image is a pointcloud of the model (~4 mil. Points), the center is the untextured model, and the right one is the textured model.
The development of the project CAPTURE makes extensive usage of our Computer Vision Development Framework ARGOS. From the software development side, it is necessary to work with large images and merge information from multiple sources simultaneously. We therefore also put special attention in parallel processing of large amount of data as well as clustering capabilities.
The application of this project is the accurate reconstruction of large scenes which includes industrial facilities, touristic and cultural heritage sites, as well as urban environments.
[/vc_column_text][/vc_column][/vc_row]