This paper introduces the Describable Textures Dataset (DTD), a large collection of 5,640 texture images collected "in the wild" and annotated with 47 descriptive texture attributes. These attributes, such as banded, cobwebbed, and zigzagged, capture a wide range of visual properties of textures. The DTD is designed to support real-world applications where texture properties are important for recognition and description. The paper addresses the challenge of automatically estimating these properties from images and demonstrates their effectiveness in texture recognition tasks. The paper proposes two texture representations: the Improved Fisher Vector (IFV), a low-level texture descriptor based on SIFT features, and the describable attributes themselves, which serve as mid-level descriptors. The IFV outperforms specialized texture descriptors on both the DTD and established material recognition datasets. When combined with IFV, the describable attributes significantly outperform the state-of-the-art by more than 8% on the FMD and KTH-TIPS-2b benchmarks. The paper also explores the use of the describable attributes for material recognition, showing that they can be used as effective texture descriptors. The attributes are easy to describe and can serve as intuitive dimensions for exploring large collections of texture patterns. The paper demonstrates that the describable attributes can be used to generate intuitive descriptions of materials and images, and that they can be combined with other features such as SIFT and color descriptors to further improve performance. The paper also addresses the practical issue of crowd-sourcing the large set of joint annotations for the DTD. The authors use Amazon Mechanical Turk to collect annotations, taking into account the co-occurrence statistics of attributes, the appearance of the textures, and the reliability of annotators. They also address the issue of noisy annotations by using multiple annotators and consensus-based methods to improve the quality of the annotations. The paper evaluates the performance of the IFV and the describable attributes on several material recognition datasets, including CUREt, UMD, UIUC, KTH-TIPS, and FMD. The results show that the IFV significantly outperforms other texture descriptors on these datasets. The describable attributes, when combined with IFV, achieve even better performance, demonstrating their effectiveness as texture descriptors. The paper concludes that the DTD and its associated methods provide a powerful tool for texture recognition and description in real-world applications.This paper introduces the Describable Textures Dataset (DTD), a large collection of 5,640 texture images collected "in the wild" and annotated with 47 descriptive texture attributes. These attributes, such as banded, cobwebbed, and zigzagged, capture a wide range of visual properties of textures. The DTD is designed to support real-world applications where texture properties are important for recognition and description. The paper addresses the challenge of automatically estimating these properties from images and demonstrates their effectiveness in texture recognition tasks. The paper proposes two texture representations: the Improved Fisher Vector (IFV), a low-level texture descriptor based on SIFT features, and the describable attributes themselves, which serve as mid-level descriptors. The IFV outperforms specialized texture descriptors on both the DTD and established material recognition datasets. When combined with IFV, the describable attributes significantly outperform the state-of-the-art by more than 8% on the FMD and KTH-TIPS-2b benchmarks. The paper also explores the use of the describable attributes for material recognition, showing that they can be used as effective texture descriptors. The attributes are easy to describe and can serve as intuitive dimensions for exploring large collections of texture patterns. The paper demonstrates that the describable attributes can be used to generate intuitive descriptions of materials and images, and that they can be combined with other features such as SIFT and color descriptors to further improve performance. The paper also addresses the practical issue of crowd-sourcing the large set of joint annotations for the DTD. The authors use Amazon Mechanical Turk to collect annotations, taking into account the co-occurrence statistics of attributes, the appearance of the textures, and the reliability of annotators. They also address the issue of noisy annotations by using multiple annotators and consensus-based methods to improve the quality of the annotations. The paper evaluates the performance of the IFV and the describable attributes on several material recognition datasets, including CUREt, UMD, UIUC, KTH-TIPS, and FMD. The results show that the IFV significantly outperforms other texture descriptors on these datasets. The describable attributes, when combined with IFV, achieve even better performance, demonstrating their effectiveness as texture descriptors. The paper concludes that the DTD and its associated methods provide a powerful tool for texture recognition and description in real-world applications.