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

Figure 7

Performance comparison of mechanisms for discounting occluded pixels. All three cases use a convolutional network of identical architecture and trained on the SORBO-combined dataset. Both mechanisms for discounting occluded pixels use the ground-truth segmentation provided with the image data. The discounting mechanism is used during both the training and testing phases. In the “none” case, the training and testing images are passed through to the classifier with no attempt to discount occlusions. In the “attenuate” case, occlusion pixels are set to black before going to the classifier. In the “inpaint” case, occlusion pixels are filled in using a digital inpainting procedure. With a convolutional network classifier, attenuation is worse than the unmodified data. Inpainting, however, is dramatically more effective than the other candidates at every level of occlusion.