Clustering Test 9 (87,800 frames, 2000BG, 1000FG)

Posted: May 2, 2013 at 3:49 pm

This is the first test showing an entire day-night cycle (approximately 24 hours). As expected, the plots show the number of percepts and processing time (segmentation + clustering) drops down to nearly nothing during the night:

87800_debug_2000bgc_1000fgc

87800_summary_2000bgc_1000fgc

The distribution of the confidence of the background percepts here is quite interesting. Also note that over this period a small amount of foreground percepts (probably 1, since we see no peak at “4.0”) were merged 4 times rather than three as in previous tests. For the background precepts in the previous test, the initial distance threshold (the distance at which to consider a percept as belonging to a cluster) was 2.3, and in this test I tightened it up to 1.9. For the foreground percepts in the previous test, the threshold was 20 and I changed it to 10. Interesting that a more strict tolerance for initial clusters lead to a greater number of merges in this test. It is unlikely the extra time made a difference, because it was at over night and very few foreground percepts were seen. With these results in mind, I’ll make the tolerance for initial clusters even more strict for the next test. Following are groups of 100 stacked percepts per image.

Foreground

outputFG-0 outputFG-1 outputFG-2 outputFG-3 outputFG-4 outputFG-5 outputFG-6 outputFG-7 outputFG-8 outputFG-9

Background

outputBG-0 outputBG-1 outputBG-2 outputBG-3 outputBG-4 outputBG-5 outputBG-6 outputBG-7 outputBG-8 outputBG-9 outputBG-10 outputBG-11 outputBG-12 outputBG-13 outputBG-14 outputBG-15 outputBG-16 outputBG-17 outputBG-18 outputBG-19 outputBG-20

Note the relative lack of a hard upper left edge in these percepts. Previously, the merging code rescaled the smaller of the two percepts to the size of the larger one, regardless of their aspect ratios, and then averaged them. The current code would only merge percepts if their aspect ratios and areas were within a particular range, and if they were to be merged, then don’t rescale but only do the weighted average. These changes thus far result in much more aesthetically interesting results. In particular, the background percepts are much softer and more interesting, due to them being composed of many more constituent parts. These percepts do still appear quite dark, which seems to indicate there are not quite enough percepts to create a good representation of all lighting conditions.

Following are two selected images (one FG and one BG) where the opacity of the percepts is determined by their degree of confidence (mean of their masks), compare to those above:

outputBG-4-meanMask outputFG-0-meanMask