channel-wise saliency

[bootchk]

An enhancement.

My initial thought was that retaining color channels throughout would give better results. In my current code I calculate an orientation difference for each channel RGB, in range [0,pi] but then sum that into one angle difference kernel (theta) in the range [0, 3pi.] Still using a kernelSum of K(d, theta). That seems ineffective e.g. for a yellow highway stripe.

So I will try again, calculating a kernel sum of K(d, thetaR, thetaG, thetaB). Do you have an intuition about that?

Consider depth-enabled photography, where the image has additional channels: infrared IR and laser-ranging depth e.g. the Intel RealSense cameras. In many cases, one channel might have the most readily available object (saliency) information. Without any apriori knowledge of which channel that is, a general-purpose algorithm (for nieve GIMP users) should work channel-wise?

[bootchk]

This morning I implemented K(d, thetaR, thetaG, thetaB) and it still doesn't recognize yellow highway stripe as salient. I will see if the stripe is salient in any channel taken singly. I am also wondering whether eliminating or lessening the weighting might help. Since as you suggest the weighting lowers the entropy of featureless areas, the yellow stripe and the road it is on are rather uniform.

Re depth-enabled, I haven't easily found such image data.

[bootchk]

I did create an RGBD image. It and its components are in my repository at data/RGBD/rgbdfromtumunichstructurenotexture.png etc.

Its a difficult test. It is no surprise that the algorithm does not find many edges in the RGB version, and does find more object edges in the RGBD version, but even the depth data does not let it find the frontmost wall corner.

The name of the file (as given by TechUMunich) indicates the nature of the image: mostly one color and texture with varied depth. The image in this post doesn't do it justice, it seems to have flattened the image taking the fourth channel as transparency. /

[kkduncan]

Sorry for the delay, but I will try to respond to all your issues / suggestions.

I'm not sure about the kernel sum of K(d, thetaR, thetaG, thetaB). I've noticed in the past when other authors presented work for color saliency, they used a different color space or just one that was in tune with the human visual system.

[kkduncan]

You may have to define saliency in a different manner in order to highlight the yellow stripe. The relational distribution based on feature differences may not be ideal in that case.

You are doing interesting work. I will give some thought to your ideas and get back to you.