Rank |
Algorithm Description |
qrmse* |
Error Hist.† |
PSNR‡ |
Time✢ | Performance∗ |
|
|
[HU] |
|
[dB] |
[s] | |
1 | ElmerSubmitter: Sam Hawker Institution: Nikon MetrologyRabbitCT dataset version: 2Date of submission: 2016-07-21
Elmer is a multi-GPU implementation using CUDA 5.5 | 0.16 |  | 88.29 | 0.3 | 100.00% |
2 | Elmer (high accuracy)Submitter: Sam Hawker Institution: Nikon MetrologyRabbitCT dataset version: 2Date of submission: 2016-07-21
High accuracy version of Elmer avoiding the fixed-point bilinear interpolation hardware | 0.05 |  | 98.47 | 0.3 | 89.47% |
3 | AccuRabbitSubmitter: Oleg Konings,Tobias Funk ,Scott Hsieh,Paul Kahn Institution: Triple Ring TechnologiesRabbitCT dataset version: 2Date of submission: 2016-07-30
We have developed a fast accurate 2 GPU implementation of the back projection algorithm for output volumes of sizes 256^3, 512^3, 1024^3 and 2048^3 | 0.03 |  | 102.18 | 0.6 | 44.74% |
4 | ThumperSubmitter: Timo Zinßer Institution: Siemens AGRabbitCT dataset version: 2Date of submission: 2013-06-11
Thumper is a CUDA-based back-projection implementation. | 0.17 |  | 87.39 | 0.6 | 43.59% |
5 | RapidRabbitSubmitter: Eric Papenhausen, Ziyi Zheng Institution: Stony Brook UniversityRabbitCT dataset version: 2Date of submission: 2013-06-04
A CUDA 3.0 based back projection implementation using a variety of optimization techniques. | 0.16 |  | 88.29 | 0.7 | 35.88% |
6 | CERA on GTX 680Submitter: Matthias Elter Institution: Siemens AGRabbitCT dataset version: 2Date of submission: 2013-02-15
The CUDA 5.0 based CERA back-projection implementation (extended with CUDA streams) running on a NVIDIA GTX 680. | 0.18 |  | 87.38 | 0.9 | 29.31% |
7 | GAPBckPrjSubmitter: Giovanni Di Domenico Institution: University of Ferrara - ItalyRabbitCT dataset version: 2Date of submission: 2016-07-29
GAPBckPrj is a CUDA 8.0 back-projection implementation. | 0.16 |  | 88.27 | 1.0 | 26.01% |
8 | CERA on GTX 670Submitter: Matthias Elter Institution: Siemens AGRabbitCT dataset version: 1 Date of submission: 2012-06-11
A CUDA 4.2 based back-projection implementation running on a NVIDIA GTX 670 GPU. | 2.84 |  | 63.17 | 3.4 | 7.73% |
9 | CERA on GTX 570Submitter: Matthias Elter Institution: Siemens AGRabbitCT dataset version: 1 Date of submission: 2012-06-11
A CUDA 4.1 based back-projection implementation running on a NVIDIA GTX 570 GPU. | 2.84 |  | 63.17 | 3.9 | 6.78% |
10 | LionEatsRabbitSubmitter: Wolfgang Wein Institution: White Lion Technologies AGRabbitCT dataset version: 1 Date of submission: 2011-08-03
The back-projection component of our rapid iterative reconstruction implemented in OpenGL / GLSL. | 2.16 |  | 65.54 | 5.9 | 4.52% |
11 | fastrabbitEXSubmitter: Jan Treibig Institution: RRZE (FAU Erlangen)RabbitCT dataset version: 1 Date of submission: 2011-07-21
OpenMP parallel implementation with SSE assembly kernel. A variant with higher precision is also available, see details. | 1.68 |  | 67.73 | 7.4 | 3.56% |
12 | CLeopatraSubmitter: Sebastian Schuberth Institution: Zuse Institute Berlin, Department Visualization and Data AnalysisRabbitCT dataset version: 1 Date of submission: 2011-01-19
This is an OpenCL-based implementation in order to evaluate its performance against CUDA, and its scalability across multiple GPUs / CPUs. | 2.86 |  | 63.12 | 10.5 | 2.52% |
13 | ConradCLSubmitter: Andreas Maier Institution: Pattern Recognition Lab, Friedrich-Alexander University Erlangen-NurembergRabbitCT dataset version: 2Date of submission: 2013-09-06
OpenCL Backprojector implemented in CONRAD. | 0.16 |  | 88.28 | 12.4 | 2.14% |
14 | SpeedyGonzalesSubmitter: Christopher Rohkohl Institution: Pattern Recognition Lab, Department of Computer Science, Friedrich-Alexander University Erlangen-Nuremberg, GermanyRabbitCT dataset version: 1 Date of submission: 2009-06-26
SpeedyGonzales is a CUDA 2.1 implementation. It uses the nice capabilities of the cuda texture-based interpolation. | 2.84 |  | 63.17 | 14.7 | 1.80% |
15 | fastrabbitSubmitter: Jan Treibig Institution: RRZE (FAU Erlangen)RabbitCT dataset version: 1 Date of submission: 2011-06-07
OpenMP parallel implementation with SSE assembly kernel. | 2.50 |  | 64.29 | 15.0 | 1.77% |
16 | RoadRunnerSubmitter: Christian Siegl Institution: Pattern Recognition Lab, Department of Computer Science, Friedrich-Alexander University Erlangen-Nuremberg, GermanyRabbitCT dataset version: 1 Date of submission: 2011-05-23
OpenCL implementation using the CUDA 3.1 framework (OpenCL 1.1). Uses texture-based interpolation and coalesced memory access. | 2.84 |  | 63.17 | 16.8 | 1.58% |
17 | TomatoSalad (formerly LolaTBBSSE)Submitter: Hannes G. Hofmann Institution: Pattern Recognition Lab, FAU ErlangenRabbitCT dataset version: 1 Date of submission: 2009-06-16
LolaBunny multi-threaded and vectorized using Intel's Threading Building Blocks (TBB) and SSE. | 0.03 |  | 101.81 | 518.9 | 0.05% |
18 | LolaSSESubmitter: Hannes G. Hofmann Institution: Pattern Recognition Lab, FAU ErlangenRabbitCT dataset version: 1 Date of submission: 2009-06-16
LolaBunny vectorized using SSE. | 0.03 |  | 101.81 | 840.7 | 0.03% |
19 | LolaTBBSubmitter: Hannes G. Hofmann Institution: Pattern Recognition Lab, FAU ErlangenRabbitCT dataset version: 1 Date of submission: 2009-06-16
LolaBunny multi-threaded using Intel's Threading Building Blocks (TBB). | 0.02 |  | 106.65 | 1,287.5 | 0.02% |
20 | LolaOMPSubmitter: Hannes G. Hofmann Institution: Pattern Recognition Lab, FAU ErlangenRabbitCT dataset version: 1 Date of submission: 2009-06-16
LolaBunny multi-threaded using OpenMP. | 0.00 |  | ∞ | 1,345.4 | 0.02% |
21 | LolaBunnySubmitter: RabbitCT Team Institution: Pattern Recognition Lab, Department of Computer Science, Friedrich-Alexander University Erlangen-Nuremberg, GermanyRabbitCT dataset version: 1 Date of submission: 2009-06-25
This is the official RabbitCT reference implementation. | 0.00 |  | ∞ | 2,509.5 | 0.01% |
Notes: |
* | Mean squared error. qrmse = sum(error2) / Nvoxels. |
† | The error histogram shows the distribution of the absolute errors. Bins 0–9 count errors between (bin, (bin+1)] HU, bin 10 counts all errors > 10 HU. Figures show relative amount: y = count / Nvoxels * 100. |
‡ | Peak SNR. PSNR = 10.0 * log10(4095.0*4095.0 / qrmse2). |
✢ | Runtime for backprojection of all views, measured in seconds. |
∗ | Performance relative to fastest implementation for this problem size. perf = tavg,fastest / tavg |
version | Between RabbitCT dataset version 1 and 2 we have changed the reference volume due to an error in the first version. Learn more |
where | Nvoxels = 5123 (on this page) and Nviews = 496 for the RabbitCT data set. |