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

Figure 4

Classification results on SORBO as a function of occlusion level and training set with the Perceptron and ConvNet algorithms. The occlusion bins partition the testing samples as defined in Figure 3. Results with the combined training set are either indistinguishable or better than the corresponding results with the unoccluded or occluded training sets with only one exception. A ConvNet training on unoccluded data outperforms the other two training options when tested on unoccluded data. Performance degrades more rapidly than the other conditions as the level of occlusion in the testing images increases, however.