COMPARE.md

Comparing with C++ Code

This code is a Python API for Video Non-Local Bayesian Denoising (VNLB). The original C++ code is from Pablo Arias. For easier interfacing, a Swig-Python Wrapper of the original C++ Code is available here.

CPU v.s. GPU Parallelism

The outputs from this VNLB code and the C++ Code are almost equal. The primary difference between to two method is the way in which we achieve parallelism. This difference impacts the final PSNR, especially on smaller images. The C++ code uses OPENMP to execute parallel CPU threads and chunks the image into sections. In this code base, a batch of random pixels are selected to searched and this searching is executed in parallel on a GPU. A secondary (and seemingly less important) difference between the two methods is how the precision of the inherent different of matrix multiplication between the C++ LAPLACK package the Python Pytorch package. However, since the image denoising community is interested in differences of approximately 1e-3 (and smaller), these small errors can potentially meaningfully change the final denoised PSNR.

Numerical Comparison

To demonstrate the difference between the two methods, we provide the compare_cpp.py script. We have two examples from the C++ Code provided in the data/ folder. For reproducibility, details to re-create the C++ Code results are included in the docs/VNLB.md file. To run the comparison, run the following:

$ python scripts/compare_cpp.py

The script prints the below table. Each element of the table is the sum of the absolute relative error between the outputs from the Python API and C++ Code.

	basic	denoised
Ave Rel. Error (basic)	0.023508	nan
cpp_psnr	31.4275	31.6715
py_psnr	31.4137	31.6513
Abs. Error (PSNR)	0.0138222	0.0201801
Rel. Error (PSNR)	0.000439814	0.000637169
Ave Rel. Error (denoised)	nan	0.0213285

Compute Time Comparison

On small images, the CPU is x10 faster than the GPU code. On larger images, the two methods are approximately equal... The Python code takes about 3 seconds longer than the C++ Code to execute.

First, export a short-cut to the directory

$ export PYVNLB_HOME="/pick/your/path/"

such that ls $PYTHON_HOME/ shows "data/" with "data/davis_baseball" as a subdirectory.

To time the C++ algorithms, please install the original C++ code. Then run the following command in the "scripts" directory,

$ time `./vnlb-gt.sh $PYVNLB_HOME/data/davis_baseball/%05d.jpg 0 4 20 $PYVNLB_HOME/data/davis_baseball_64x64/vnlb/ "-px1 7 -pt1 2 -px2 7 -pt2 2 -verbose"`

#davis
real	2m25.276s
user	12m25.894s
sys	3m5.471s

$ cd vnlb/
$ rm -r ./__pycache__
$ time `python ./scripts/example.py`

#davis
real	2m28.202s
user	12m18.676s
sys	3m10.941s

On the davis example, the original execution time of the C++ Code and Python API is X and Y, respectively. ...