Seite - 120 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure 3. Two examples of active contour modeling. Crops of car images (left), the inferred segmentation masks (middle) and contour segments (right) before ac- tivecontourmodeling (red)andafter (blue). Inferred segmentation masks are thresholded and the resulting sharp separating contour between the foreground and background region is subjected to a refinement procedure. The contour is split into cer- tain and uncertain regions depending on the neural networks certainty in its prediction. A contour re- gion is considered to be certain if nearby values in the corresponding segmentation mask are close to 0 or1, anduncertainotherwise. Uncertain Regions These contour regions typi- cally occur in over- or underexposed areas of an im- age and are iteratively adjusted using active contour modeling [7]. The method aims to minimize an en- ergy functional of a spline contour in the inferred segmentation mask. Figure 3 shows the effects of the approach on two examples. The first example showsitspositiveinfluencewhile thesecondisafail- ure case. CertainRegions Cars typically have large regions that are smooth for aerodynamic and aesthetic rea- sons. Edges that are present however can be rather sharp. The motivation of the following procedure, which we call adaptive smoothing, is to mimic this bias. We aim to perform a high degree of smooth- ing without displacing the contour by more than0.5 pixels. The upper limit is enforce since based on the neuralnetworkassessmentsuchasegment isalready close to the ground truth target. As an initial step the contour segment is split into separate sequences forxandy coordinates. The fol- lowing procedure is applied separately to both. Let κ= (κi) N i=1 be such a sequence of real points. We use Gaussian filtersGσi with standard deviationsσi thatadept to thecurrentposition. AkernelGσi isob- 1 2 3 4 5 1 2 3 4 5 Figure 4. Comparison of a contour before postprocessing (red) and after adaptive smoothing (blue). Full segmenta- tion mask and contours (bottom second from the left) and fiveenlarged regions. tained by sampling a Gaussian density in the points Z∩ [−2σi,2σi]andnormalizing. Thesmoothedcontourκshasequalshapetoκand isdefinedas κsi := ( κ∗Gσi ) i . (6) For the computation of the valuesσiwe are looking for the largest kernel that displaces κ less than 0.5 pixels. A naive implementation of this idea has two issues: First the set { σi∈R≥0 : ∣∣κi−κsi∣∣<0.5} might not be bounded and second this approach can lead to large jumps in consecutive entries ofσ. For this reason we pose the definition with additional re- strictions: (B) σ1=σN=0, (C) |σi−σi+1|≤α, i∈{1. ..N−1}, (M) σi∈R≥0maximal s.t. |κi−κsi|<0.5. Under these conditions solutions exist and are unique. Requirement (B) enforces the fixed bound- ary conditions κ1 = κs1 and κN = κ s N while (C) ensures continuity within the contour segment. The parameterα specifies an upper bound for the slope. In practice the settingα= 0.5 performs well. For the implementation of this method it is advisable to only consider a discrete set of possible values forσi. A comparison of contours before and after postpro- cessingcanbeseen inFigure4. 5.Conclusion We studied methods for the generation of highly accurate binary segmentation masks, including weightingschemesthat improvedtheperformanceof default loss functions and a novel Gradient Loss. In addition we developed a specialized postprocessing procedure that exploits a bias in our dataset. We cre- ated a solution that poses a significant upgrade over 120