Triplet dissimilarity: improving dissimilarity approaches through metric learning

The dissimilarity representation describes each sample by its differences from other samples. Earlier work uses traditional metric distances, which have some limitations. Metric learning can improve this by learning a task-specific metric. We join the two ideas, a CNN backbone trained with triplet loss learns a task-specific dissimilarity function and feeds it to two representations: Triplet Dissimilarity Space (TDS), which measures every image against a set of prototypes, and Triplet Dissimilarity Vector (TDV), which turns each image-prototype pair into its own training point. We test on three fine-grained image datasets: CUB-200-2011 birds, Stanford Cars, and FGVC-Aircraft, using three embedding sizes and three clustering methods.