3D Reconstruction meets GPGPU meets Image Analysis

The main focus of my work is addressing 3D reconstruction of environments from multiple images and video sequences. I report on the experiences gained during our work on an automated, large scale urban reconstruction system, and I describe our specific Bayesian approach to infer the correctness of image-based two-view geometries.

A significant part of my presentation is dedicated to the efficient generation of high-quality, dense 3D models from images. Our continuous approach is robust against noise and outliers found in depth maps e.g. generated by fast local stereo, and it is highly suitable for acceleration by programmable graphics processing units. The complete pipeline from images to dense models (consisting of dense stereo, optional clean-up of depth maps, and robust depth map fusion) is significantly sped up by GPUs without sacrificing the quality of resulting models.

Finally, I briefly sketch our continuous, GPU-accelerated approaches for low-level image analysis tasks like segmentation.