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

Figure 8

Inpainting improves performance on both occluded and unoccluded test images. The “control” and “occluded training” conditions are the unoccluded and combined training conditions from Figure 6. The “recovery” condition is the inpainting result from Figure 7. The “recovery” condition is consistently better as occlusion increases. Discounting occluded pixels using inpainting has the unexpected benefit of also increasing performance on unoccluded images. This is a dataset augmentation effect. The training set in the “recovery” and “occluded training” conditions is ten times larger than in the “control” condition. Inpainting allows the network to leverage this larger training set without overfitting.