Image from a daVinci intervention and reconstructed point cloud.
Outline of the surface reconstruction workflow.
Comparison between the structure of the cpu and gpu version of the HRM: (a) cpu HRM structure; (b) cpu HRM recursion over whole image; (c) gpu HRM structure; and (d) gpu HRM recursion over overlapping sub images.
gpu consistency check and disparity correction procedure.
Hybrid cpu-gpu algorithm workflow.
Simulation environment: (a) image from virtual camera; (b) disparity image; (c) overlay of virtual liver with reconstructed surface; and (d) side view of reconstructed surface.. [URL: http://dx.doi.org/10.1118/1.3681017.1]10.1118/1.3681017.1
Selected frames with different textures: cpu, gpu, cpu-gpu, and both opencv surface reconstruction disparity images and disparity difference (>1 pixel) to ground truth disparities.
Heart 1 sequence: (a) frame 5; (b) disparity image; (c) Reconstructed point cloud; and (d) Color-coded distance map (from green (error = 0 mm) to red (error = 2 mm)) between reconstructed surface and ground truth: blue = no comparable points found.
Parameterization of the evaluation with virtual image sequences.
Evaluation of the cpu, gpu, cpu-gpu, the cpu- and gpu-opencv surface reconstruction on nine image sequences from the simulation environment; best results in bold print.
Errors due to two error sources: Gaussian noise and poorly rectified images.
Mean distance between reconstructed and ground truth point cloud for two image sequences of the silicone phantom and the cpu version, the cpu-gpu version and both opencv algorithms.
Run time comparison in ms between the cpu, the gpu, and the cpu-gpu version of the algorithms as well as both opencv correspondence analysis algorithms for left and right image with resolution of 320 × 240, 640 × 480, and 960 × 540.
Article metrics loading...
Full text loading...