Image Pre-processing vs. Transfer Learning for Visual Route Navigation

This paper investigates image pre-processing and triplet learning for place recognition in route navigation. The first contribution combines image pre-processing and ImageNet pre-trained neural networks for generating improved image descriptors. The second contribution is a fast, compact ‘FullDrop’ layer that can be appended to an ImageNet pre-trained network and taught to generate invariant image descriptors with triplet learning. The proposals decrease inference time by 8x and parameters by 30x while keeping comparable performance to NetVLAD, the state of the art for this task