TPSNet: Reverse Thinking of Thin Plate Splines for Arbitrary Shape Scene Text Representation

Wang, Wei; Zhou, Yu; Lv, Jiahao; Wu, Dayan; Zhao, Guoqing; Jiang, Ning; Wang, Weipinng

doi:10.1145/3503161.3547882

Cited by 14 publications

(6 citation statements)

References 47 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…TextDCT (Su et al 2022) transforms text instance masks to the Frequency domain by discrete cosine transform (DCT), and then extracts the low-frequency components to represent the text instance masks. TPSNet (Wang et al 2022a) utilizes thin plate splines (TPS) to parameterize text contours into TPS fiducial points.…”

Section: Regression-based Methodsmentioning

confidence: 99%

“…In this paper, we use LRA to compactly represent text contours. Unlike previous parameterized text shape methods that use curve fitting (Liu et al 2020;Wang et al 2022a) or mask compression (Su et al 2022), LRA is a data-driven approach that represents text boundaries in a low-dimensional space by exploiting the distribution of labeled text contours.…”

Section: Methodology Low-rank Approximation Representationmentioning

confidence: 99%

“…It adopts a dense assignment branch to provide ample supervised signals for training, and meanwhile, accelerates the inference speed through a sparse assignment branch, enjoying the merits of both branches. Specifically, following previous regression-based text detection network Wang et al 2022a), our dense assignment scheme treats the text region as the positive sample region. In our sparse assignment scheme, we construct a prediction-aware matrix S for selecting K positive samples for each text instance, The matrix element is defined as…”

Section: Dual Assignment Schemementioning

confidence: 99%

“…Some methods adopt contour points or parametric curves to fit the text shape, but they either lack imposing geometric constraints or fail to consider (a) TextRay (Wang et al 2020a) (b) TextDCT (Su et al 2022) (c) FCENet (d) Ours The other problem is that regression-based methods often overlook the overall speed of the entire pipeline. Specifically, current regression-based methods can be divided into dense assignment approaches (Liu et al 2020;Su et al 2022;Wang et al 2022a) and one-to-one assignment approaches (Raisi, Younes, and Zelek 2022;Ye et al 2023;Shao et al 2023) according to the allocation of positive samples. However, dense assignment approaches require nonmaximum suppression (NMS) to filter a large number of redundant predictions, and this process is time-consuming, especially for arbitrary-shaped redundant predictions.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

LRANet: Towards Accurate and Efficient Scene Text Detection with Low-Rank Approximation Network

Su,

Chen,

Shao

et al. 2024

AAAI

View full text Add to dashboard Cite

Recently, regression-based methods, which predict parameterized text shapes for text localization, have gained popularity in scene text detection. However, the existing parameterized text shape methods still have limitations in modeling arbitrary-shaped texts due to ignoring the utilization of text-specific shape information. Moreover, the time consumption of the entire pipeline has been largely overlooked, leading to a suboptimal overall inference speed. To address these issues, we first propose a novel parameterized text shape method based on low-rank approximation. Unlike other shape representation methods that employ data-irrelevant parameterization, our approach utilizes singular value decomposition and reconstructs the text shape using a few eigenvectors learned from labeled text contours. By exploring the shape correlation among different text contours, our method achieves consistency, compactness, simplicity, and robustness in shape representation. Next, we propose a dual assignment scheme for speed acceleration. It adopts a sparse assignment branch to accelerate the inference speed, and meanwhile, provides ample supervised signals for training through a dense assignment branch. Building upon these designs, we implement an accurate and efficient arbitrary-shaped text detector named LRANet. Extensive experiments are conducted on several challenging benchmarks, demonstrating the superior accuracy and efficiency of LRANet compared to state-of-the-art methods. Code is available at: https://github.com/ychensu/LRANet.git

show abstract

Section: Regression-based Methodsmentioning

confidence: 99%

Section: Methodology Low-rank Approximation Representationmentioning

confidence: 99%

Section: Dual Assignment Schemementioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

LRANet: Towards Accurate and Efficient Scene Text Detection with Low-Rank Approximation Network

Su,

Chen,

Shao

et al. 2024

AAAI

View full text Add to dashboard Cite

show abstract

“…To investigate the effectiveness of the dual matching method, we compare it with sparse matching-only and dense matching. To make a fair comparison with dense matching, we adopt the best current dense matching strategy in [82]. As shown in Table III, compared with the sparse matching-only scheme, our dual matching scheme is able to provide more supervised signals to facilitate the learning of classification and regression features, which in turn achieves a 1.1% F-measure surpassing.…”

Section: Alation Study 1) Different Positive Sample Matching Methodmentioning

confidence: 99%

An efficient scene text detection neural network

2022

The 6th International Conference on Control Engineering and Artificial Intelligence

View full text Add to dashboard Cite

Recently, regression-based methods, which predict parameter curves for localizing texts, are popular in scene text detection. However, these methods struggle to balance concise structure and fast post-processing, and the existing parameter curves are still not ideal for modeling arbitrary-shaped texts, leading to a challenge in balancing speed and accuracy. To tackle these challenges, we firstly propose a dual matching scheme for positive samples, which accelerates inference speed through sparse matching scheme and accelerates model convergence through dense matching scheme. Then, we propose a novel text contour representation method based on low-rank approximation by exploiting the shape correlation between different text contours, which is complete, compact, simplicity and robustness. Based on these designs, we implement an efficient and accurate arbitrary-shaped text detector, named LRANet. Extensive experiments are conducted on three challenging datasets, which demonstrate the accuracy and efficiency of our LRANet over state-of-the-art methods. The code will be released soon.

show abstract