Table 1 A comparison of CUDA-based stereo matching algorithms on Middlebury and KITTI datasets
From: Real-time CUDA-based stereo matching using Cyclops2 algorithm
Method | Middlebury (1436 × 992 single image) | KITTI (1242 × 375 single image) | Advantages/disadvantages | ||
|---|---|---|---|---|---|
Avg. runtime, s | Bad pixels (error > 4 pixels), % | Avg. runtime, s | Bad pixels (error > 3 pixels), % | ||
BM [5] | x | x | 0.1 | 25.27 | Sparse results, high error rates, does not work in untextured areas, CPU implementation |
SGM [6] | 9.9 | 12.2 | x | x | Non optimal matching cost, performance depends on parameter choice, CPU implementation |
SGBM [8] | 2.27 | 18.5 | 1.1 | 10.86 | Simple matching cost, fast computation time even with CPU implementation |
CSCT [10] | x | x | 0.0064 | 8.24 | Simple GPU model, hard to adjust to complex tasks |
IDR [11] | 0.49 | 12.7 | x | x | Strict binary weights near boundaries |
MC-CNN-ACRT [12] | 150 | 4.48 | 67 | 3.89 | Complexity of training CNN, slow speed |
MC-CNN-FST [12] | 1.69 | 6.7 | x | x | CNN optimised for speed, requires powerful GPU |
C-CNN [13] | x | x | 1.00 | 4.54 | Smoothing technology dependent |
DNET [14] | x | x | 0.06 | 4.34 | Dataset specific, added complexity of CNN |