Figure 2.

Some representative failure cases on the ScanNetv2 [16] dataset. The first row shows that the baseline model lost morphological integrity and accuracy in shade, leading to an extra part of the trash can’s structure. The second row shows that Fusenet confused part of a single item as other objects while adjacent pixels of the same object have evident contextual relationships in the depth image. The third row shows that both the prediction by the baseline model and the ground truth neglected the clothes on the bed.