In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009 
Figure 6. First pair of pictures: User clicks per screen required to obtain correct road mapping subject to number of image 
parts already seen by our leamable detection algorithm (a) - road], road2 data (b) - road3 and road4 data. Solid squares 
show necessary number of clicks if our detection algorithm is applied before user input. Empty squares show necessary 
number of clicks without use of automatic detection algorithm. Y-axis shows estimated number of clicks and x-axis 
represents number of processed images. Second pair of pictures: Error of automatic detection algorithm subject to number of 
image parts already seen by leamable detection algorithm (c) - roadl, road data (d) - road3 and road4 data. The error is 
measured as a fraction of image square misclassified by our detection algorithm. At all plots green lines correspond to roadl 
and blue lines correspond to road2. Y-axis shows error rate and x-axis represents number of processed images. 
are about 0.5 megapixel size and correspond to 5-10 meters 
of road surface. Some of image parts with results of our 
automatic detection are shown in Figure 8. 
Experiments setup 
We have developed a testing framework which emulates user 
activity at the on-line stage. Given the classification results 
and ground truth data it starts with automatic thresholds 
adjusting. Gradient descent algorithm is used to determine a 
set of thresholds that minimizes total area of misclassified 
objects. Then user interaction is emulated as follows. At first, 
our framework corrects all errors of automatic detection 
which can be amended by relabeling segments of the coarsest 
segmentation scale. Then testing framework emulates user- 
aided error correction at subsequent segmentation scale. This 
procedure is repeated up to the most detailed segmentation 
scale. Total number of clicks required for errors correction is 
calculated as a sum of click counts at all segmentation scales. 
This statistic measures overall usability of our tool for road 
mapping. 
One can see total clicks count per image part measured on 4 
roads from our image base in Figure 6 (a, b). Roadl and 
road2 contain greater number of road defects than road3 and 
road4, therefore larger number of user clicks is required for 
mapping last two roads. We have compared number of clicks 
required to achieve accurate mapping when user corrects 
errors of our automatic detection algorithm with the number 
of clicks required for mapping road surface from scratch 
when no automatic detection is performed. One (?- или It can 
be seen) can see that usage of automatic detection algorithm 
leads to advance in usability of mapping tool. 
As a matter of fact, road marking can be usually found 
perfectly after processing the second or the third part image. 
So, the problem of road defects detection is more 
challenging. Figure 6 (c, d) demonstrates misclassified area 
of road defects subject to number of image parts seen by 
detection algorithm. 
Figure 7 illustrates false positive and false negative error 
rates of road defect by pixels on roadl data. This picture 
represents usual behaviour of our system. The rate of 
detected defects increases over time when more defects 
examples shown to automatic detection algorithm. 
In summary, overall error tends to decrease while the number 
of handled images grows. The system usually starts to 
distinguish road defects since two or three images have been 
handled. Some road images like road3 and road4 contain a 
small amount of road defects (some image parts do not 
contain them at all). Although learning process is slowed 
down and benefit of using interactive system is reduced on 
such kind of roads however, usage of automatic detection 
result still remains beneficial. 
8. CONCLUSIONS AND FUTURE WORK 
We have presented a tool for efficient interactive mapping of 
road defects and lane marking on rectified images of road 
pavement surface. Intensive use of computer vision methods 
on different stages of our data processing workflow increases 
usability of the tool. 
The most significant drawbacks of our tool is the limitation 
of using segments in user interaction stage and incapability to 
correct detection results on sub-segment level. Also our 
system currently is unable to accommodate to changes of the 
road structure, e.g. illumination level changing. This 
drawback can be eliminated if we provide on-line classifier 
with concept adapting. 
Figure 7. False positive and false negative rates on roadl 
data subject to a number of handled road sections. Y-axis 
shows error rate and x-axis represents number of processed 
images.