COMPARE.md

Comparing with C++ Code

This code is a Python API for Video Non-Local Bayesian Denoising (C++ code originally from Pablo Arias). The numerical outputs from the Python API and the C++ Code are exactly equal. The Python code takes about 3 seconds longer than the C++ Code to execute.

Numerical Comparison

To demonstrate this claim, 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 with the provided C++ outputs, type:

$ export OMP_NUM_THREADS=4
$ python compare_cpp.py

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

	noisyForFlow	noisyForVnlb	fflow	bflow	basic	denoised
Total Error (cv2)	0.000505755	0	504.308	21.643	0	0
Total Error (cpp)	0	0	0	0	0	0

The following describes each column:

• noisyForFlow: the images used to compute the optical flow
• noisyForVnlb: the images to be denoised
• fflow: the forward optical flow (t -> t+1)
• bflow: the backward optical flow (t -> t-1)
• basic: the "basic" esimate, as described in the VNLM method
• denoised: the denoised images

The graphic below depicts the input-output relationship of the columns:

noisyForFlow -> (fflow, bflow)
(noisyForVnlb, fflow, bflow) -> (basic, denoised)  

In the above table, the two rows identify two methods used to read image data. The first row uses opencv (cv2) and the second row uses the original, wrapped C++ functions (cpp). Images for optical flow (noisyForFlow) are read with the iio library. Images for the VNLB method (noisyForVnlb) are read with the VidUtils library.

For optical flow, images read with opencv are slightly different from the images read using the C++ function, as indicated by the 0.0005 under the noisyForFlow column. This yields a change in the the optical flow outputs (fflow and bflow). However, this small change in optical flow yields no difference in the final denoising results (basic and denoising).

For the VNLB method, images read with opencv are exactly equal the images read using the C++ function, as indicated by the 0 in the noisyForVnlb column.

Compute Time Comparison

The Python code takes about 3 seconds longer than the C++ Code to execute. To time the algorithms, one can execute both methods within a time bracket.

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

#davis_64x64
real	0m5.917s
user	0m30.099s
sys	0m7.698s

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

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

#davis_64x64 
real	0m8.867s
user	0m41.212s
sys	0m10.648s

#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 2:25 and 2:28, respectively. This increase in time is from an increase in execution time within the C++ routines themselves, rather than the Python wrapper. See the scripts/example.py and the runVnlbTimed function for more information.