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

Figure 9

Analysis of segmentation errors. The automated segmentation algorithm leverages the planar structure of the occlusions to estimate which pixels are occlusion and which pixels are figure. Each pixel falls into one of four conditions. In the “hit” and “correct rejection” cases, the estimate is correct. A hit occurs when the algorithm predicts the presence of an occlusion and an occlusion is actually present. A correct rejection occurs when the algorithm accurately predicts the lack of an occlusion. The “miss” and “false alarm” cases are both errors. A miss occurs when the algorithm predicts the lack of an occlusion, but an occlusion is present. A false alarm occurs when the algorithm predicts an occlusion, but no occlusion is present. The accuracy of the automated segmentation process varies depending on the level of occlusion in the image pair. For unoccluded image pairs, the process is entirely accurate. At higher levels of occlusion, overall accuracy drops. False alarms, however, are much more common than misses. This indicates a bias towards marking a given pixel as an occlusion.