One-Shot Re-identification using Image Projections in Deep Triplet Convolutional Network
Gábor Kertész, Imre Felde · 2020
Representation learning of images using deep neural networks have shown great results in classificational tasks. In case of instance recognition, or object re-identification other approaches are used. Siamese architectured convolutional networks were the first approach to learn from semantic distances, and give the similarity of two inputs. Triplet networks apply the triplet loss based on the furthest positive and the closest negative pair. In this paper we present a method to apply multi-directional image projections as an initial transformation to compress image data, whereafter the discriminative ability remains. After performing the training on vehicle images, the model is evaluated by measuring the one-shot classification accuracy.