Cross-dataset Learning for Generalizable Land Use Scene Classification

Dimitri Gominski (1) , Valérie Gouet-Brunet (2) , Liming Chen (3, 4)

1 UCPH - University of Copenhagen = Københavns Universitet
2 LaSTIG - Laboratoire sciences et technologies de l'information géographique
3 LIRIS - Laboratoire d'InfoRmatique en Image et Systèmes d'information
4 imagine - Extraction de Caractéristiques et Identification

Few-shot and cross-domain land use scene classification methods propose solutions to classify unseen classes or un-seen visual distributions, but are hardly applicable to real-world situations due to restrictive assumptions. Few-shot methods involve episodic training on restrictive training subsets with small feature extractors, while cross-domain methods are only applied to common classes. The underlying challenge remains open: can we accurately classify new scenes on new datasets? In this paper, we propose a new framework for few-shot, cross-domain classification. Our retrieval-inspired approach 1 exploits the interrelations in both the training and testing data to output class labels using compact descriptors. Results show that our method can accurately produce land-use predictions on unseen datasets and unseen classes, going beyond the traditional few-shot or cross-domain formulation, and allowing cross-dataset training

Domaines: Vision par ordinateur et reconnaissance de formes [cs.CV]; Intelligence artificielle [cs.AI]; Apprentissage [cs.LG]; Traitement du signal et de l'image [eess.SP]

Dimitri Gominski, Valérie Gouet-Brunet, Liming Chen. Cross-dataset Learning for Generalizable Land Use Scene Classification. EarthVision'22 workshop in conjuction with the Computer Vision and Pattern Recognition (CVPR) 2022 Conference, Jun 2022, La Nouvelle Orléans, LA, United States. pp.1381-1390, ⟨10.1109/CVPRW56347.2022.00144⟩. ⟨hal-03685079⟩