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

Figure 11

Occlusion recovery using attenuation and inpainting. Sample images from the SORBO dataset (a) are recovered using either attenuation (b) or inpainting (c). These recovery methods are candidate techniques for discounting irrelevant information from the images. Attenuation removes the occluding texture by setting occluding pixels to black. Inpainting infers a plausible value for the occluded pixels by using the remaining visible pixels. The dots along the borders of the inpainted images are an artifact of a boundary condition bug in the inpainting library routine.

(a) Sample images from SORBO dataset

(b) Occluded regions discounted by attenuation

(c) Occluded regions discounted by inpainting