This website requires JavaScript.

ELFIS: Expert Learning for Fine-grained Image Recognition Using Subsets

Pablo VillacortaJes\'us M. Rodr\'iguez-de-VeraMarc Bola\~nosIgnacio Saras\'uaBhalaji NagarajanPetia Radeva
Mar 2023
Fine-Grained Visual Recognition (FGVR) tackles the problem of distinguishinghighly similar categories. One of the main approaches to FGVR, namely subsetlearning, tries to leverage information from existing class taxonomies toimprove the performance of deep neural networks. However, these methods rely onthe existence of handcrafted hierarchies that are not necessarily optimal forthe models. In this paper, we propose ELFIS, an expert learning framework forFGVR that clusters categories of the dataset into meta-categories using bothdataset-inherent lexical and model-specific information. A set of neuralnetworks-based experts are trained focusing on the meta-categories and areintegrated into a multi-task framework. Extensive experimentation showsimprovements in the SoTA FGVR benchmarks of up to +1.3% of accuracy using bothCNNs and transformer-based networks. Overall, the obtained results evidencethat ELFIS can be applied on top of any classification model, enabling theobtention of SoTA results. The source code will be made public soon.