|Description (include details on usage, files and paper references)||Textile retrieval in real environments is a poorly investigated research field besides fashion cloths retrieval. Up to our knowledge, there is no publicly available dataset that focuses on the recognition of rigid and non-rigid textiles presented in different sizes, shapes and capturing conditions. For this reason, we created a new dataset for the retrieval of textiles in bedrooms.
The dataset is composed of 684 images of sizes that range between 480x360 and 1280x720 pixels obtained from 15 videos of YouTube. The videos were recorded in bedrooms with plenty of textiles and with different camera poses, illumination conditions, occlusions, etc., which makes the textile retrieval task very challenging. The dataset contains 67 classes of textiles such as curtains, carpets, sofas, shirts, dresses, etc. In one image, several classes of textiles may appear. The number of elements of each class varies from 4 to 116. There is a total of 1913 regions. Therefore, the dataset is highly skewed, simulating a real scenario. All images can be found in the folder "All".
We labelled the dataset in order to provide a ground truth that allows the user to automatically evaluate the performance of a method on the dataset. The ground truth includes the bounding box coordinates and the class labels of each textile region in the images of the dataset. We provide the ground truth in the form of an XML file. The folder "GroundTruth" is composed of 67 subfolders, one subfolder per class, that contain all the image regions for each class.
TextilTube dataset can be very interesting in fields like child sexual abuse or robbery to connect evidences of different investigations and also for marketing studies in textile stores to suggest the products which best fit the decoration of users rooms