Skip to main content

Table 7 Average error improvement after applying suggested post-processing, i.e. median blur filter. Post-processing significantly improved produced results. The added computation time was not significant. We performed the tests using only our proposed algorithms on the two Middlebury training sets images

From: Real-time CUDA-based stereo matching using Cyclops2 algorithm

Algorithm Improvement on average error, % Latency, s
TX1 TK1
Quarter resolution dataset
 MD 22.95 0.004 0.013
 RMA 12.20 0.003 0.013
 MNS1 7.41 0.005 0.015
 MNS2 15.79 0.004 0.018
 MNS3 15.79 0.004 0.011
 Average 14.59 0.004 0.014
Half resolution dataset
 MD 24.59 0.040 0.306
 RMA 16.67 0.293
 MNS4 34.21 0.036 0.307
 MNS5 10.53 0.039 0.308
 Average 25.16 0.038 0.303