Toward Fast Computation of Dense Image Correspondence on the GPU
Large-scale video processing systems are needed to support human analysis of massive collections of image streams. Video, from both current small-format and future large-format camera systems, constitutes the single largest data source of the near future, dwarfing the output of all other data sources combined. A critical component to further advances in the processing and analysis of such video streams is the ability to register successive video frames into a common coordinate system at the pixel level. This capability enables further downstream processing, such as background/mover segmentation, 3D model extraction, and compression. We present here our recent work on computing these correspondences. We employ coarse-to-fine hierarchical approach, matching pixels from the domain of a source image to the domain of a target image at successively higher resolutions. Our diamond-style image hierarchy, with total pixel counts increasing by only a factor of two at each level, improves the prediction quality as we advance from level to level, and reduces potential grid artifacts in the results. We demonstrate the quality our approach on real aerial city imagery. We find that registration accuracy is generally on the order of one quarter of a pixel. We also benchmark the fundamental processing kernels on the GPU to show the promise of the approach for real-time video processing applications.
- Research Organization:
- Lawrence Livermore National Lab. (LLNL), Livermore, CA (United States)
- Sponsoring Organization:
- USDOE
- DOE Contract Number:
- W-7405-ENG-48
- OSTI ID:
- 924953
- Report Number(s):
- UCRL-CONF-233816; TRN: US200809%%613
- Resource Relation:
- Conference: Presented at: 2007 High Performance Embedded Computing (HPEC) Workshop, Lexington, MA, United States, Sep 18 - Sep 20, 2007
- Country of Publication:
- United States
- Language:
- English
Similar Records
3-D Semantic Information Inference from Airborne Video. Final Report
A Public-Private-Academic Partnership to Advance Solar Power Forecasting