A Model for Simulating User Interaction in Hierarchical Segmentation
SUPPLEMENTARY MATERIAL

Content

Segmentation tasks

In this section we describe the segmentation tasks considered in the experiments. Each of the 20 segmentation tasks is composed by an input image, a textual description explaining which objects should be segmented, and the ground truth.

TaskInput imageGround truthTask descriptionImage credits
1Segment each of the four wings. Split each wing into 2 pieces, separating the central region with another color. Courtesy of contributor
2Segment each cell. Courtesy of contributor
3Segment each of the 6 strawberries. Include the sepals (leaves) and peduncles (cables). Do not include the shadows in the background. Multispectral Image Database
4Segment each of the four planes of the image (sky and 3 mountains). McGill Calibrated Colour Image Database
5Segment the regions of water, grass and sky, separating them from the other regions. Multispectral Image Database
6Segment the squares from the texture. Keep the rest (background) in a single region. McGill Calibrated Colour Image Database
7Segment each of the balloons. Keep the two nozzles that appear in image attached to their balloons. Do not include pieces of the background. Multispectral Image Database
8Segment each cell. Broad Bioimage Benchmark Collection - BBBC01
9Segment the letters "T", "i", "d" and "e" of the largest package. Segment separately the dot of the "i" and the holes in the letters "d" and "e". Caltech - Home Objects dataset
10Segment the man and each lantern. Consider only the yellow part with inscription of each lantern (do not include the red part). The Berkeley Segmentation Dataset and Benchmark
11Segment each of the 10 mirrors. McGill Calibrated Colour Image Database
12Segment each of the peppers. Include the peduncles (cables). Do not include the shadows in the background. Multispectral Image Database
13Segment the polyhedron, separating individually each visible face. HIPR2 Image Library
14Segment the mushroom. The Berkeley Segmentation Dataset and Benchmark
15Segment the flower, separating individually each one of the petals and the stamens region (central part). McGill Calibrated Colour Image Database
16Segment each of the balls. Do not include the shadows in the background. Multispectral Image Database
17Segment each of the worms. Do not include pieces of the background. HIPR2 Image Library
18Segment the flower, merging all the petals in a single region. Do not include pieces of the background. McGill Calibrated Colour Image Database
19Segment each of the spools. Multispectral Image Database
20Segment the image into three regions: the orange isolated rightmost, the other oranges in the same region and the rest of the image in another region. McGill Calibrated Colour Image Database

Testing policy 3 with several probability values

In this section we present charts of the experiments when using distinct constant values for P in policy 3.

Below, we show the correlation between human profiles and the ones from the proposed policies and also for all the P values tested, as shown in the previous charts.