Total-Text: A Comprehensive Dataset for Scene Text Detection and Recognition

I. INTRODUCTION

• Existing datasets offer limited curved-text coverage: ICDAR texts are horizontal, MSRA-TD500 texts are mostly straight, and CUTE80 has only 80 images with minimal scene diversity.Source: §I · p. 1
• Total-Text contains 1,555 scene images and 9,330 annotated words spanning horizontal, multi-oriented, and curved text.The dataset was collected with curved text in mind to fill a gap in scene-text datasets.Source: §I · p. 1
• Recent methods –[11] cast text detection as semantic segmentation and report state-of-the-art results on ICDAR’11, ’13, and MSRA-TD500, as well as successful curved-text detection.The paper also examines how He et al.’s system, which assumes no orientation and uses heuristic grouping, handles multiple text orientations in natural scenes.Source: §I · pp. 1–2

II. RELATED WORKS

• The related-works section covers scene-text datasets and detection systems, and recommends for a fuller discussion.Source: §II · p. 2

A. Scene Text Datasets

• ICDAR’03 had 509 horizontal-text images; ICDAR’11 reduced this to 484 to remove duplication, and ICDAR’13 further trimmed it to 462 while expanding categories and tasks.Source: A. Scene Text Datasets · p. 2
• ICDAR’15 [13] used 1,670 wearable-device images, with arbitrary-orientation text that was often out of focus.Source: A. Scene Text Datasets · p. 2
• MSRA-TD500 addressed the shortage of arbitrarily oriented text with 300 training and 200 testing images annotated using minimum-area rectangles.Source: A. Scene Text Datasets · p. 2
• COCO-text [14] contains 63,686 images and 173,589 labeled regions across horizontal, arbitrary, and curved orientations, but uses axis-oriented rectangles.The paper notes that this ground truth seems applicable only to horizontal and vertical text.Source: A. Scene Text Datasets · p. 2
• CUTE80 was the only publicly available curved-text dataset to the authors’ knowledge, with just 80 images and limited sceneries.Source: A. Scene Text Datasets · p. 2

B. Scene Text Detection:

• Earlier SWT and MSER methods detected text by grouping or extracting components, then using component properties and classifiers.Both represented characters better than color, edge, and texture feature extractors, and component-based methods typically generated text lines, filtered candidates, and segmented them.Source: B. Scene Text Detection · p. 3
• CCTN cascades supervised FCNs for text-region and text-line detection, generalizes across orientations, and performs strongly on ICDAR 2013 and MSRA-TD500.The passage describes it as one of the best-performing systems on both datasets.Source: B. Scene Text Detection · p. 3
• Zhang et al. trained a Text-Block FCN for local and global text-region context and a Character-Centroid FCN to filter false text-line candidates.They argued that character detection’s local nature restricted CNN potential.Source: B. Scene Text Detection · p. 3
• CNN component classifiers improved feature extraction robustness and distinguished text from non-text,, demonstrating strong discriminative power.Huang et al. applied a character classifier to MSER-generated components.Source: B. Scene Text Detection · p. 3
• Text-line generation remained a key challenge: methods, often handcrafted it, while He et al. [10] trained an FCN to infer text-line candidates.The passage notes that robust CNNs had not substantially benefited the grouping of characters into words.Source: B. Scene Text Detection · p. 3

III. TOTAL-TEXT DATASET

• The Total-Text dataset section covers its collection motivation, horizontal, multi-oriented, and curved text, algorithm orientation assumptions, and dataset statistics.The passage outlines these as the section’s topics.Source: §III · p. 3

Dataset attributes

• 4,265 of Total-Text’s 9,330 text instances are curved, addressing a gap in existing scene-text datasets.The paper notes that ICDAR datasets contain only horizontal text and MSRA-TD500 has just two curved-text instances.Source: A. Dataset Attributes · pp. 3–4
• Curved words vary in character orientation, challenging straight-line and consistent-orientation assumptions in multi-oriented detectors,.The paper reports that these algorithms and failed on the same curved-text images in MSRA-TD500; that dataset contains only two curved instances.Source: A. Dataset Attributes · pp. 3–4
• Total-Text’s images combine diverse orientations, including horizontal, multi-oriented, curved, circular, and wavy text.More than half of the images contain combinations of more than two orientations, as illustrated by the listed two- and three-orientation examples.Source: Figure 6 · p. 4, A. Dataset Attributes · p. 4
• The authors favor focused scene-text images as a starting point for curved-text research, despite newer datasets’ emphasis on incidental scenes.They argue that focused datasets such as ICDAR and MSRA-TD500 helped advance the field, while incidental images add complexity.Source: A. Dataset Attributes · p. 4
• Tightly fitted polygon annotations bind word-level text regions while avoiding the excess background included by rectangular boxes.The dataset also provides rectangular boxes, recognition groundtruth with spatial coordinates, orientation labels, binary masks, and do-not-care labels for non-English, watermarked, or unreadable text. DetEval is adapted to polygon groundtruth.Source: A. Dataset Attributes · pp. 4–5, Figure 5 · p. 4

B. Dataset Statistics

• The dataset has 1,255 training images and 300 testing images; comparisons use ICDAR 2013 and MSRA-TD500 as common horizontal and multi-oriented benchmarks.Source: B. Dataset Statistics · p. 5
• Total-Text contains 9,330 annotated texts, averages 6 instances per image and 1.8 orientations per image, and ranks first against competitors on both numbers.More than half of its images contain two or more orientations.Source: B. Dataset Statistics · p. 5
• About half of text instances are curved, with the remainder split nearly equally between horizontal and multi-oriented text; curved forms include horizontal, vertical, circular, and wavy.Curved text usually appears alongside horizontal or multi-oriented text, mixing orientations within images and challenging detection robustness and generalization.Source: B. Dataset Statistics · pp. 5–6
• Unlike CUTE80, whose images are mostly football jerseys, Total-Text covers diverse scenes: business locations account for 61.2% and tourist spots 21.1% of curved-text instances.Source: B. Dataset Statistics · p. 6

• The authors fine-tuned DeconvNet and evaluated it on Total-Text.Source: §IV · p. 6
• FCN-inspired methods cast text detection as segmentation and achieved state-of-the-art results –[11].Source: §IV · p. 6
• Although conventional algorithms often failed on curved text, segmentation methods succeeded in limited examples because benchmarks were scarce.Source: §IV · p. 6
• The authors were particularly motivated by [10], which achieved good results without the heuristic grouping rules used by most other algorithms.Source: §IV · p. 6

A. DeconvNet

• DeconvNet was selected for its Pascal VOC semantic-segmentation results and its multiple deconvolutional layers, which allow fine-grained observation of deviations.The paper notes that its scope is not to propose a new curved-text solution; it converts and fine-tunes the network for text localization.Source: A. DeconvNet · p. 6
• The conversion replaced DeconvNet’s 21 output layers for 20 PASCAL VOC classes and background with two layers for text and non-text, then fine-tuned the pre-trained model in one step.Other training implementations remained consistent with the original paper, apart from the conversion and training data.Source: A. DeconvNet · p. 6
• Training used legible COCO-text data [14], converted to binary text masks and cropped into patches, with patches below 10% text removed.About 15k images yielded roughly 200k training and 80k validation patches; augmentation used horizontal flipping and random cropping to 224x224.Source: A. DeconvNet · p. 6

B. Experiments

• Inference resizes images to 224×224, thresholds the saliency map at 0.5, then groups text pixels with connected components and tightly bounds them with polygons.Source: B. Experiments · p. 7
• DeconvNet’s Total-Text evaluation was unsatisfactory, with failures on patterned backgrounds and when multiple word candidates were merged.The evaluation is listed in Table I; the passages do not provide its cell values.Source: Table I · p. 7, B. Experiments · p. 7
• Feature-map inspection found activations at lower layers and finer details in later layers, with characters becoming recognizable as processing proceeded.The network produced candidates across orientations without orientation assumptions or heuristic grouping.Source: B. Experiments · p. 7
• DeconvNet’s unpooling and learnable upsampling produce high-resolution saliency maps depicting text shape and orientation, requiring minimal post-processing.The passage relates spatially preserved feature maps to pixelwise prediction in and [10].Source: B. Experiments · p. 7
• Low-confidence curved-text boundaries may benefit from text-line supervision; [10] reports F-score falling from 0.84 to 0.5 without FTN.Figure 12 illustrates examples where the network is less confident about curved-text region shapes.Source: B. Experiments · p. 7, Figure 12 · p. 7

V. CONCLUSION

• Total-Text introduces curved text as a missing element in scene-text datasets and treats it as part of the multi-oriented text detection problem.The paper hopes the dataset’s availability will advance research on curved text.Source: §V · pp. 7–8
• DeconvNet analysis suggests spatial resolution and contextual information are crucial to segmentation-based methods, which predict text regions across orientations without hard-coded rules.Source: §V · p. 8
• The authors plan to explore segmentation further toward designing a scene-text detector effective against multi-oriented text.Source: §V · p. 8

VI. APPENDIX

• Total-Text examples include perspective distortion and variation in font type and size.These are among the challenging attributes illustrated in Figure 13.Source: §VI · p. 9
• Other illustrated challenges include text-like backgrounds, uneven lighting, and low contrast between text and background.Examples of text-like backgrounds include bricks and trees.Source: §VI · p. 9

A. Text Orientations

• The orientation examples range from unified orientation to two orientations and then to images containing all sorts of orientations.Figure 14 illustrates these examples.Source: A. Text Orientations · p. 9, Figure 14 · p. 10

B. Different Groundtruth Evaluations

• The evaluation compares precision and recall for DeconvNet detections against two groundtruth formats, shown as red rectangles and blue polygons.Detections are shown in green.Source: B. Different Groundtruth Evaluations · pp. 9–10
• The polygon-shaped groundtruth tightly bounds text regions and provides a more accurate evaluation than conventional rectangular groundtruth.Table II presents the evaluation with the two groundtruth formats.Source: B. Different Groundtruth Evaluations · p. 9, Table II · p. 10
• Missed large intersections with groundtruth regions lower both recall and precision in the illustrated detections.The examples also show higher precision with rectangular groundtruth for some detections.Source: B. Different Groundtruth Evaluations · p. 10

C. Groundtruth Examples

• Each image is annotated with text-instance locations, transcripts, orientations, and binary masks marking annotated regions as white and background as black.The paper says Figures 16 and 17 depict annotation details.Source: C. Groundtruth Examples · p. 10
• Polygon-shaped groundtruth provides a more accurate representation of algorithm performance than rectangle-shaped groundtruth.The figure illustrates disagreements between polygon- and rectangle-shaped regions.Source: Figure 15 · p. 11
• Polygon-shaped bounding boxes tightly bound text, and the annotations are more comprehensive.Source: Figure 16 · p. 12