[Paper Review] Detecting Curve Text in the Wild: New Dataset and New Solution

Yuliang Liu, Lianwen Jin+2 more|arXiv (Cornell University)|Dec 6, 2017

Handwritten Text Recognition TechniquesComputer Science26 references212 citations

TL;DR

The paper introduces CTW1500, a curve-text dataset, and a polygon-based Curve Text Detector (CTD) with a recurrent offset connection (TLOC) that directly detects curve text and outperforms prior methods."

ABSTRACT

Scene text detection has been made great progress in recent years. The detection manners are evolving from axis-aligned rectangle to rotated rectangle and further to quadrangle. However, current datasets contain very little curve text, which can be widely observed in scene images such as signboard, product name and so on. To raise the concerns of reading curve text in the wild, in this paper, we construct a curve text dataset named CTW1500, which includes over 10k text annotations in 1,500 images (1000 for training and 500 for testing). Based on this dataset, we pioneering propose a polygon based curve text detector (CTD) which can directly detect curve text without empirical combination. Moreover, by seamlessly integrating the recurrent transverse and longitudinal offset connection (TLOC), the proposed method can be end-to-end trainable to learn the inherent connection among the position offsets. This allows the CTD to explore context information instead of predicting points independently, resulting in more smooth and accurate detection. We also propose two simple but effective post-processing methods named non-polygon suppress (NPS) and polygonal non-maximum suppression (PNMS) to further improve the detection accuracy. Furthermore, the proposed approach in this paper is designed in an universal manner, which can also be trained with rectangular or quadrilateral bounding boxes without extra efforts. Experimental results on CTW-1500 demonstrate our method with only a light backbone can outperform state-of-the-art methods with a large margin. By evaluating only in the curve or non-curve subset, the CTD + TLOC can still achieve the best results. Code is available at https://github.com/Yuliang-Liu/Curve-Text-Detector.

Motivation & Objective

• Motivate reading curve-shaped text in the wild and address dataset scarcity.
• Create CTW1500, a curve-text dataset with polygon annotations to ease labeling.
• Propose a direct polygon-based detector (CTD) that detects curved text without post-hoc grouping.
• Introduce TLOC to capture sequential context between point offsets for smoother localization.
• Provide universal training that adapts to rectangular, quadrilateral, or curve annotations.

Proposed method

• Propose CTD, a polygon-based text detector that regresses 14 curve points plus circumscribed rectangle parameters.
• Split regression into width and height offsets to improve training stability.
• Attach a recurrent transverse and longitudinal offset connection (TLOC) using BLSTM to model sequential context among point offsets.
• Utilize PSROIPooling to generate per-point offset features and feed them into the TLOC module for sequential prediction.
• Train with a multi-task loss combining classification and localization terms across proposals.
• Apply two post-processing steps: non-polygon suppression (NPS) and polygonal non-maximum suppression (PNMS).

Experimental results

Research questions

• RQ1 Can a polygon-based detector directly localize curved text without assembling components from separate detections?
• RQ2 Does incorporating a recurrent offset connection (TLOC) improve localization accuracy for curve text?
• RQ3 How do curve-focused post-processing steps (NPS, PNMS) affect detection performance on curve and non-curve text?
• RQ4 Is the approach universal enough to handle curve, rectangular, and quadrilateral text annotations without extra labeling effort?

Key findings

• CTW1500 contains 1,500 images with 10,751 bounding boxes, including 3,530 curve boxes.
• CTD (without TLOC) achieves state-of-the-art Hmean of 69.5 on the full CTW1500 test set; CTD with TLOC reaches 73.4.
• CTD+TLOC improves curve-text detection substantially, achieving 77.1 R, 57.1 P, 65.6 H on the curve subset with 13.3 FPS.
• PNMS consistently improves over NMS across experiments; adding TLOC yields the largest gains in Hmean (~4 percentage points).
• CTD+TLOC outperforms several baselines (SegLink, SWT, CTPN, EAST, DMPNet) on the CTW1500 results, especially in the curve subset (e.g., ~28% Hmean improvement).
• The method remains fast (13–15 FPS) while handling curve text directly and generalizes to non-curve text subsets.