Sensitivity of visual attention to coding artifacts on still pictures

O. Le Meur
Part III, Chapter 1, section 1.5 of Le Meur's Phd thesis (unpublished results)

Home Main idea Degree of similarity between human saliency maps stemming from an original and distorted pictures

Main Idea

The goal of the experiment conducted in 2005 was to examine the sensitivity of our visual attention to coding artifacts. Eye tracking experiments were conducted in free-vieiwng tasks and involved up to 40 observers. The degradation was obtained by either a JPEG or JPEG2000 coding. In the following, the similarity between human saliency maps is presented.

Degree of similarity between human saliency maps stemming from an original and distorted pictures

Click on the pictures to look at the original picture.

Qualitative analysis based on fixation maps:

• Figure 1.2 (page 147):

Barba and Isabe pictures (on the left) were encoded by a JPEG (middle) and a JPEG2000 (right) encoder. Compression artifacts are visible.






• Figure 1.3 (page 148):

Yellow crosses correspond to a fixation point. The rows illustrates from left to right the fixation maps for the original, the JPEG and the JPEG200 pictures. The first row shows fixations occuring between 0 and 2 seconds of viewing. The second row shows fixations between 0 and 14 seconds of viewing.

Linear correlation between fixation maps of distorted and original pictures:

• Table 1.1 and 1.2 (page 147):



Linear correlation in function of the viewing time. Correlation between original picture and its distorted version (JPEG).

Pictures / Viewing time	2s	8s	14s
Barba	0.986	0.979	0.981
Isabe	0.988	0.991	0.987




Linear correlation in function of the viewing time. Correlation between original picture and its distorted version (JPEG2000).

Pictures / Viewing time	2s	8s	14s
Barba	0.981	0.984	0.98
Isabe	0.991	0.979	0.979

From the qualitative and quantitative analysis, our preliminary conclusions was that our visual attention is not significantly modified in presence of coding artifacts. However, it was difficult to conclude definitively since we only tested two kinds of degradations and two pictures (that are maybe not appropriate due to the presence of faces).