Deep metric learning with distance sensitive entangled triplet losses

Download
2021-2-12
Karaman, Kaan
Metric learning aims to define a distance that is able to measure the semantic difference between the instances in a dataset. The most recent approaches in this area mostly utilize deep neural networks as their models to map the input data into a feature space by finding appropriate distance metrics between the features. A number of loss functions are already defined in the literature based on these similarity metrics to discriminate instances in the feature space. In this thesis, we particularly focus on triplet loss functions in order to designate their gradients. It is argued that the gradients of the vanilla triplet loss function do not force the instances in a triplet along the right direction with the right magnitude. Hence, the similarities between the instances in a triplet and the natural phenomena of a free electrostatic charge being affected by several forces due to the other charged bodies located in certain coordinates in the space are exploited to determine the right direction and magnitude. Considering the partial gradients of the loss function with respect to the anchor, positive and negative instances of any valid triplet generated from the dataset, four novel triplet loss functions are proposed that cope with the problem pointed out. It is shown that these loss gradients gradually solve the drawbacks of the vanilla loss function. The performance increment of these losses, especially the METU loss, over the other triplet losses is presented by the results of several fair experiments on a commonly used fine-grained dataset: CUB200-2011. The results of the proposed techniques are comparable with respect to the score values of the state-of-the-art methods in the deep metric learning topic.
Citation Formats
K. Karaman, “Deep metric learning with distance sensitive entangled triplet losses,” M.S. - Master of Science, Middle East Technical University, 2021.