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

Figure 3

Classification results on SORBO as a function of occlusion level and training set with the chance algorithm. Points outside the box plots indicate outliers. The chance classification algorithm ignores the training data and makes a random guess for each testing sample. SORBO contains five classes with an approximately equal number of examples in each. The expected chance performance is therefore 20%. The occlusion bin percentage indicates the upper bound, inclusive, of the bin. For example, the 0% bin contains testing samples with zero occluded pixels and the 27% bin contains samples with greater than 0% and less than or equal to 27% occluded pixels. As expected, the chance algorithm scores an accuracy of approximately 20% for all trials, at every occlusion level and with every variant of the SORBO training set.