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

Heavily occluded objects are more difficult for classification algorithms to identify correctly than unoccluded objects. This effect is rare and thus hard to measure with datasets like ImageNet and PASCAL VOC, however, owing to biases in human-generated image pose selection. We introduce a dataset that emphasizes occlusion and additions to a standard convolutional neural network aimed at increasing invariance to occlusion. An unmodified convolutional neural network trained and tested on the new dataset rapidly degrades to chance-level accuracy as occlusion increases. Training with occluded data slows this decline but still yields poor performance with high occlusion. Integrating novel preprocessing stages to segment the input and inpaint occlusions is an effective mitigation. A convolutional network so modified is nearly as effective with more than 81% of pixels occluded as it is with no occlusion. Such a network is also more accurate on unoccluded images than an otherwise identical network that has been trained with only unoccluded images. These results depend on successful segmentation. The occlusions in our dataset are deliberately easy to segment from the figure and background. Achieving similar results on a more challenging dataset would require finding a method to split figure, background, and occluding pixels in the input.