Total-Text is a benchmark dataset for scene text detection, introduced in 2017 by researchers at the University of Toronto and other institutions. It was designed to address the limitations of earlier datasets that primarily focused on horizontal or roughly aligned text, by including a significant proportion of curved and arbitrarily oriented text instances. The dataset has become a standard evaluation resource for machine learning and deep learning models in the field of scene text detection, a subfield of artificial intelligence and computer vision.
The dataset consists of 1,555 images collected from real-world scenes, such as street signs, storefronts, and posters. Each image is annotated with word-level bounding polygons that precisely outline the text regions, including curved and multi-oriented text. The annotations are provided in a format that supports both polygon and rectangular representations, enabling flexible evaluation of detection algorithms. Total-Text includes a training set of 1,255 images and a test set of 300 images, with a total of over 9,000 annotated text instances. The dataset is notable for its high proportion of curved text, which poses a significant challenge for traditional detection methods that assume axis-aligned or horizontal text.
Total-Text has been widely used in academic research and competitions, such as the ICDAR (International Conference on Document Analysis and Recognition) challenges, to benchmark the performance of text detection systems. It has driven the development of advanced neural network architectures, including those based on residual networks and U-Net-like segmentation approaches, as well as methods that leverage data augmentation and batch normalization to improve robustness.
Dataset Composition and Annotation
The images in Total-Text are sourced from a variety of urban environments, capturing text in different languages, fonts, lighting conditions, and perspectives. The annotations are performed at the word level, with each word enclosed by a polygon that follows the text's shape, whether curved, rotated, or horizontal. The dataset also provides a separate set of axis-aligned bounding boxes for each word, which can be used for evaluation methods that require rectangular regions. The ground truth is stored in a text file format, where each line corresponds to a word and includes the polygon coordinates and the transcription of the text.
Evaluation Metrics and Protocols
Standard evaluation for Total-Text follows the protocols established by the ICDAR competitions, using precision, recall, and F-measure as primary metrics. Detection results are matched to ground truth polygons based on intersection-over-union (IoU) thresholds, typically set at 0.5. The dataset's official evaluation code is available to ensure consistent comparison across methods. Researchers are encouraged to report results on the test set, and many published papers include comparisons on Total-Text alongside other datasets like ICDAR 2015 and MSRA-TD500.
Impact on Scene Text Detection Research
Total-Text has significantly influenced the design of modern text detection systems. Early methods relied on hand-crafted features and connected component analysis, but the complexity of curved text prompted a shift toward deep learning approaches. Many state-of-the-art models now employ fully convolutional networks, multi-head attention mechanisms, and encoder-decoder architectures to predict text regions in a pixel-wise manner. The dataset has also been used to train and evaluate end-to-end text spotting systems that combine detection and recognition, often leveraging sequence-to-sequence models and transformer backbones.
Limitations and Extensions
While Total-Text provides a challenging benchmark, it has limitations, including a relatively small number of images compared to some modern datasets and a focus on English text. Subsequent datasets, such as CTW1500 and ArT, have expanded the scope to include more diverse text shapes and languages. Nevertheless, Total-Text remains a widely cited and used resource for evaluating the robustness of text detection algorithms, particularly in handling curved and arbitrarily oriented text.