This paper introduces FGVC-Aircraft, a new dataset containing 10,000 images of aircraft spanning 100 aircraft models, organized in a three-level hierarchy. The dataset includes 100 variants, 70 families, and 30 manufacturers. The images are annotated with the model and bounding box of the dominant aircraft. The dataset was constructed using images collected from aircraft enthusiasts and online resources. The images were filtered to maximize internal diversity and reduce unwanted correlations. Bounding boxes were crowdsourced using Amazon Mechanical Turk, and hierarchical labels were manually inspected. The dataset is publicly available for research purposes only. The paper presents baseline results on aircraft model identification, showing that the performance is quite good for some distinctive categories but worse for others, such as Airbus or Boeing families. The accuracy for variant classification is 58.48%, and for manufacturer classification is 71.30%. The dataset is intended to introduce aircraft recognition as a novel domain in fine-grained visual classification (FGVC) to the wider computer vision community. The dataset is part of the ImageNet 2013 FGVC challenge. The images were obtained from aircraft spotter collections to maximize internal diversity. The authors plan to substantially increase the size of the dataset in the future by including more models as more photographers provide permission to use their photos. The dataset is made publicly available for research purposes only. The authors acknowledge the contributions of the photographers and the support from various organizations.This paper introduces FGVC-Aircraft, a new dataset containing 10,000 images of aircraft spanning 100 aircraft models, organized in a three-level hierarchy. The dataset includes 100 variants, 70 families, and 30 manufacturers. The images are annotated with the model and bounding box of the dominant aircraft. The dataset was constructed using images collected from aircraft enthusiasts and online resources. The images were filtered to maximize internal diversity and reduce unwanted correlations. Bounding boxes were crowdsourced using Amazon Mechanical Turk, and hierarchical labels were manually inspected. The dataset is publicly available for research purposes only. The paper presents baseline results on aircraft model identification, showing that the performance is quite good for some distinctive categories but worse for others, such as Airbus or Boeing families. The accuracy for variant classification is 58.48%, and for manufacturer classification is 71.30%. The dataset is intended to introduce aircraft recognition as a novel domain in fine-grained visual classification (FGVC) to the wider computer vision community. The dataset is part of the ImageNet 2013 FGVC challenge. The images were obtained from aircraft spotter collections to maximize internal diversity. The authors plan to substantially increase the size of the dataset in the future by including more models as more photographers provide permission to use their photos. The dataset is made publicly available for research purposes only. The authors acknowledge the contributions of the photographers and the support from various organizations.