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

Figure 2

Inpainting over occlusions. The raw input (a) is a frame of video masked with an image of retinal veins. The algorithm does not have access to the pixels covered by the mask. A fast linear diffusion solver (b) uses the raw input and a segmentation estimate to fill in over the occluded regions. Magnifying the inpainted output (c) reveals the weaknesses of this simple technique. The dirt to the left of the image contains little edge information and inpainting works well. The boundaries of the rider, however, are badly blurred where an occlusion is present. More sophisticated inpainting algorithms do not have this problem but at a cost of significant additional complexity and a much longer compute time per frame.

(a)

(b)

(c)