I found PyTorch an excellent tool to prototype CUDA-based kernels for experimenting with computer vision algorithms. I revamp my old stereo matching code with CUDA in this repository and rewrite them using modern CUDA and PyTorch.
Here's a video from the old implementation:
and a sample estimation:
Methods
- Cost measurements:
- Sum of squared distances (SSD)
- Birchfield
- Aggregation mechanisms:
- Semiglobal matching
- Disparity reducer
- Winner takes all
- Dynamic Programming
Note for AI practitioners: The operations aren't differentiable.
- Version 0.1.*: Just porting/rewriting of the old code.
- Version 1.0.*: Fix/improve unit testing and continuous integration.
- Version 1..: Optimizations of the CUDA code.
- Version 2.0.0: Experiments with features extracted from neural networks.
Requirements:
- Common C++ build tools
- CUDA Toolkit>=10.0
- Pytorch>=1.0
Installing:
$ pip install git+https://github.com/otaviog/stereo-matching---------------------------------------------------------------------------- benchmark 'aggregation': 2 tests ----------------------------------------------------------------------------
Name (time in ms) Min Max Mean StdDev Median IQR Outliers OPS Rounds Iterations
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_bm_sgm_gpu 6.8900 (1.0) 7.5300 (1.0) 6.9917 (1.0) 0.0848 (1.0) 6.9755 (1.0) 0.0391 (1.0) 5;6 143.0258 (1.0) 94 1
test_bm_sgm_cpu 451.0489 (65.46) 465.8883 (61.87) 458.6411 (65.60) 5.7751 (68.10) 457.5974 (65.60) 8.5925 (219.51) 2;0 2.1804 (0.02) 5 1
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------------ benchmark 'cost': 5 tests ------------------------------------------------------------------------------------------------
Name (time in us) Min Max Mean StdDev Median IQR Outliers OPS Rounds Iterations
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_bm_birchfield_gpu 7.2490 (1.0) 5,706.3670 (1.87) 2,987.4853 (1.20) 545.6754 (9.60) 3,024.4600 (1.21) 21.9982 (1.0) 2454;4818 334.7297 (0.84) 35469 1
test_bm_ssd_gpu 2,433.9830 (335.77) 3,057.3600 (1.0) 2,494.7826 (1.0) 56.8682 (1.0) 2,492.1500 (1.0) 25.3140 (1.15) 15;18 400.8365 (1.0) 171 1
test_bm_ssd_texture 7,141.7370 (985.20) 8,070.7620 (2.64) 7,292.0231 (2.92) 120.6339 (2.12) 7,305.9345 (2.93) 128.0390 (5.82) 24;5 137.1362 (0.34) 114 1
test_bm_upload_ssd_texture 7,322.2750 (>1000.0) 8,142.2920 (2.66) 7,497.1437 (3.01) 107.4651 (1.89) 7,508.3075 (3.01) 92.1450 (4.19) 22;4 133.3841 (0.33) 132 1
test_bm_birchfield_cpu 726,389.1040 (>1000.0) 737,151.8250 (241.11) 729,505.7956 (292.41) 4,387.7641 (77.16) 727,860.8590 (292.06) 4,178.8813 (189.96) 1;0 1.3708 (0.00) 5 1
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
------------------------------------------------------------------------------------ benchmark 'disparity_reduce': 2 tests -------------------------------------------------------------------------------------
Name (time in us) Min Max Mean StdDev Median IQR Outliers OPS Rounds Iterations
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_bm_dynprog_gpu 683.8870 (1.0) 782.7430 (1.0) 708.1196 (1.0) 16.9596 (1.0) 715.8820 (1.0) 30.0320 (1.0) 92;4 1,412.1908 (1.0) 262 1
test_bm_dynprog_cpu 39,436.6220 (57.67) 50,944.3990 (65.08) 40,664.7475 (57.43) 2,517.0496 (148.41) 39,843.0405 (55.66) 672.7905 (22.40) 2;3 24.5913 (0.02) 24 1
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------