Mitigation of Effects of Occlusion on Object Recognition with Deep Neural Networks through Low-Level Image Completion

Figure 10

Performance with inferred versus ground-truth segmentation. The “control,” “occluded training,” and “recovery, data” conditions are the same as in Figure 9. The “recovery, auto” condition is new and matches the “recovery, data” condition except for the source of the segmentation masks. The data case uses the ground-truth segmentations provided with the dataset. The automatic case uses only the raw stereo image pairs and infers the segmentation masks. These two recovery conditions perform at parity at higher levels of occlusion. Using inferred segmentations erases the dataset augmentation effect observed in Figure 9, however. Performance on unoccluded test images is no better than the occluded training condition.