Figure 1.
Tradeoff between effort and quality of segmentation.
The quality of reconstruction is shown as a function of the reconstruction effort. Randomly finding places to verify leads to slow improvements in segmentation quality.
Figure 2.
The workflow consists of boundary prediction and agglomeration including automatic and manual effort.
Figure 3.
Challenges in using edge probability to determine agglomeration strategies and manual reconstruction priority.
(a) Examining edges solely based on edge probability is not always ideal since it does not account of the size of the connected superpixels. (b) a and c should not be connected even though both local connections a–b and b–c indicate a connection.
Figure 4.
Manual verification methodology.
GPR is used to both prioritize edges to be examined and terminate the manual work.
Figure 5.
EM data samples used in the experiments.
Table 1.
Reducing reconstruction effort through prioritization.
Figure 6.
Adjusted Rand Index as a function of manual reconstruction effort.
Effort is determined by the number of yes/no decisions. GPR Priority dominates Random Priority in each sample.
Figure 7.
Shows the predictiveness of GPR similarity measure compared to the actual adjusted Rand index computed with an expert-created ground truth.
The y-axis gives similarity as a function of different agglomeration thresholds. The agglomeration is based on the mean intensity of the pixel-wise boundary prediction.
Figure 8.
Show the predictiveness of GPR similarity measure compared to the actual adjusted Rand index.
Unlike Figure 7, the agglomeration is based on the edge connection probability, not the mean intensity of the boundary.
Figure 9.
The poor predictiveness of GPR similarity measure compared to the actual adjusted Rand index.
While the agglomeration is based on mean intensity as in Figure 7, the quality of the edge probabilities is degraded by only using mean intensity as a feature for computing edge probability, thus leading to poor correspondence between GPR and the adjusted Rand index.
Figure 10.
The quality of GPR approximation as a function of the number of paths considered.
Note that small magnitude difference between one path and the maximum number of paths.