Monday, January 05, 2009

Feng-GUI attention heatmap accuracy

We’ve been asked several times about the accuracy of Feng-GUI attention heatmap and why is it not reaching 100%.
Feng-GUI attention heatmap algorithm, which we call the ViewFinder , is a composition of algorithms aggregating over 30 years of research. Research from the fields of neuro-science studies, visual attention, statistics of eye-tracking sessions, perception and cognition of humans.
We improve the ViewFinder algorithm in order to achieve the highest accuracy, but is it possible to get to 100% ?

Let us answer by asking a question:
How accurate is the "Magic Wand" in Photoshop? (also referred as the “Fuzzy select” in Gimp)

The Magic Wand tool is designed to select areas of the image by finding edges of image segments.

Part of the answer to accuracy can be found at the "image segmentation and boundary detection" benchmark conducted by Berkley.
It’s a 1000 photos test suite for algorithms that try to find the boundaries of elements inside pictures.
In the test suite results you can see how real people and computer algorithms, marked the boundaries of segments inside photos and how accurate they are.
The current score is:
Real people: 79%
Machines algorithms: 70%

original image

edge detection algorithm

edge detection by human

You can see that edge detection algorithms are not 100% accurate and will never be considering the hardware used be personal computers.
To solve this limitation, GIMP supports six different selection tools. In order of appearance in the tool box they are: rectangle, ellipse, free, fuzzy, bezier, and intelligent.
Each tool has its advantages and disadvantages--no one tool is clearly superior to the others. In general, a combination of tools proves to be the optimal strategy for complex selection tasks.

We conducted a comparison benchmark between Feng-GUI results and eye-tracking results of 40 people, 90 images, 15 seconds per image.
Comparing the regions of interests (ROI) in the images, results with more than 70% of similarities.
If you compare Feng-GUI results to random groups of 10 people (out of the 40 people), for 5 seconds, the similarities go over 80%.

To enable designers and advertisers integrate Feng-GUI’s attention heatmap into their work process, we keep on improving the ViewFinder algorithm while maintaining two important requirements: Accuracy as well as Speed.

happy new year!