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

Figure 12

Accuracy of automatic stereo segmentation. NORB and SORBO are stereo datasets. Sample images from SORBO (a) contain various types and level of occlusions. SORBO also includes corresponding ground-truth segmentation masks (b). In these masks, white identifies occlusion pixels and black identifies target pixels. The stereo estimation algorithm produces a mask using only the raw input images (a). These estimates tend to have a high false positive rate but miss occlusions only at the borders (c). In these error images, light green corresponds to a hit. Dark green is a correct rejection. Blue is a false prediction of occlusion or false positive. Red is a missed prediction.

(a) Left/right images

(b) Ground-truth masks

(c) Inferred mask error