Inconsistency Between Code Output and Paper Results

[Leozyc-waseda]

Hello,
Thank you very much for your code.

While testing your code, I've noticed that the final results differ from those presented in itti's paper. Could this be due to a need for specific parameter adjustments? If so, could you please guide me on which parameters should be altered? Below are the results I obtained:

• papers results
  

• your results
  

[Leozyc-waseda]

@akisatok san

I also noticed that in this issue, you mentioned that you changed the value of Gabor. I tried to access Professor Itti's official website to download the source code, but failed (it seems I don't have access). Could you share the original parameters?
#5 (comment)

[akisatok]

@Leozyc-waseda Sorry for the too-late reply.

My code in this repo is a re-implementation based on the paper [Itti+ TPAMI1998] (NOTE: This is not my paper.), which implies that my implementation may produce different outputs from the ones by the official implementation http://ilab.usc.edu/toolkit/.

For example, I have hard-coded several parameters, such as the size of local windows for taking local maxima, coefficients of Gabor filters, and feature weights, as presented in pySaliencyMapDefs.py. Although I have not checked those parameters of the official implementation, they might be different from mine.

[Leozyc-waseda]

@akisatok

Thank you for getting back to me. I noticed that your code used a bilateral filter instead of a Gaussian filter, as compared to the original MATLAB version.

https://github.com/TsotsosLab/SMILER/blob/427ce783a4477db550dd8fe064a38e3b71db346e/models/matlab/GBVS/makeGBVSParams.m#L23-L26

pySaliencyMap/pySaliencyMap.py

Lines 252 to 256 in b7aee8a

	# normalize
	normalizedSM = self.SMRangeNormalize(SMMat)
	normalizedSM2 = normalizedSM.astype(np.float32)
	smoothedSM = cv2.bilateralFilter(normalizedSM2, 7, 3, 1.55)
	self.SM = cv2.resize(smoothedSM, (width,height), interpolation=cv2.INTER_NEAREST)

Could you please explain the reason for this change? I have experimented with a Gaussian filter and found a significant difference in results compared to the bilateral filter.

I'm also having difficulty understanding why your final output includes both a binarized saliency map and a salient region. Due to the influence of the Otsu method and the GrabCut algorithm, the results of these two algorithms differ somewhat from the previous saliency algorithm results. What is the primary purpose of the binarized saliency map and salient region?

[akisatok]

@Leozyc-waseda Thanks for the comment. As you know, bilateral filters can blur an image while respecting strong edges. This characteristic is significant for salient region extraction, implemented as SMGetBinarizedSM

pySaliencyMap/pySaliencyMap.py

Line 260 in b7aee8a

def SMGetBinarizedSM(self, src):

and SMGetSalientRegion

pySaliencyMap/pySaliencyMap.py

Line 270 in b7aee8a

def SMGetSalientRegion(self, src):

.

The paper by Itti et al. does not introduce any smoothing filters for obtaining the final saliency map. This means that if you want to replicate the results of that paper, you can simply remove Line 255 rather than replace a bilateral filter with a Gaussian filter.

SMGetBinarizedSM and SMGetSalientRegion are just optional. You can just ignore them if you do not need them.

[Leozyc-waseda]

@akisatok
Thank you for your reply. I will remove the smoothing filters and see what results I get.

I've noticed that the results from SMGetSalientRegion effectively reflect human bottom-up attention patterns, such as traffic signs and moving vehicles, which I find very intriguing. My original intention in asking was to inquire whether this was also your intent when you implemented it.