{"notes":[{"id":"text-spotting/abcnet","category":"text-spotting","legacyPath":"text-spotting/2002-abcnet","legacySlug":"2002-abcnet","legacyPaths":[],"title":"ABCNet","subtitle":"Greetings from Bezier Curves","author":"Z. Yuan","paperDate":"2020-02-24","publishedOn":"2024-10-20","paperTitle":"ABCNet: Real-time Scene Text Spotting with Adaptive Bezier-Curve Network","paperUrl":"https://arxiv.org/abs/2002.10200","bibliography":{"title":"ABCNet: Real-time Scene Text Spotting with Adaptive Bezier-Curve Network","authors":["Yuliang Liu","Hao Chen","Chunhua Shen","Tong He","Lianwen Jin","Liangwei Wang"],"url":"https://arxiv.org/abs/2002.10200","pdfUrl":"https://arxiv.org/pdf/2002.10200","codeUrl":null,"preprint":{"id":"2002.10200","submittedOn":"2020-02-24","url":"https://arxiv.org/abs/2002.10200"},"publication":{"venue":"CVPR","year":2020,"kind":"conference","url":"https://arxiv.org/abs/2002.10200"},"citation":"@misc{textspottingabcnet2020,\n  title = {{ABCNet: Real-time Scene Text Spotting with Adaptive Bezier-Curve Network}},\n  author = {Yuliang Liu and Hao Chen and Chunhua Shen and Tong He and Lianwen Jin and Liangwei Wang},\n  year = {2020},\n  eprint = {2002.10200},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2002.10200}\n}\n","citationPath":"/content/papers/citations/text-spotting/abcnet.b93e438b307c.bib"},"thumbnail":{"src":"/content/papers/assets/0d8681d7dcbb702c346101a5dc93b5f4ee4f05dcca5293b3ec7f5962c8cf6e0d.jpg","width":1940,"height":474},"locale":"en","path":"/content/papers/en/text-spotting/abcnet.c76be2defed5.md","headings":[{"id":"defining-the-problem","text":"Defining the Problem","depth":2},{"id":"bezier-curves","text":"Bezier Curves","depth":2},{"id":"solution-overview","text":"Solution Overview","depth":2},{"id":"model-architecture","text":"Model Architecture","depth":3},{"id":"generating-groundtruth","text":"Generating GroundTruth","depth":3},{"id":"synthetic-data","text":"Synthetic Data","depth":3},{"id":"bezier-alignment","text":"Bezier Alignment","depth":3},{"id":"text-recognition-head","text":"Text Recognition Head","depth":3},{"id":"training-configuration","text":"Training Configuration","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"bezieralign-analysis","text":"BezierAlign Analysis","depth":3},{"id":"beziercurve-analysis","text":"BezierCurve Analysis","depth":3},{"id":"experimental-results-on-total-text","text":"Experimental Results on Total-Text","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"text-spotting/mask-textspotter","relation":"background","topic":"Joint text detection and recognition"},{"id":"text-spotting/abcnet-v2","relation":"next","topic":"Joint text detection and recognition"}]},{"id":"text-spotting/mask-textspotter","category":"text-spotting","legacyPath":"text-spotting/1807-mask-textspotter","legacySlug":"1807-mask-textspotter","legacyPaths":[],"title":"Mask TextSpotter","subtitle":"Localization as Recognition","author":"Z. Yuan","paperDate":"2018-07-06","publishedOn":"2024-10-21","paperTitle":"Mask TextSpotter: An End-to-End Trainable Neural Network for Spotting Text with Arbitrary Shapes","paperUrl":"https://arxiv.org/abs/1807.02242","bibliography":{"title":"Mask TextSpotter: An End-to-End Trainable Neural Network for Spotting Text with Arbitrary Shapes","authors":["Pengyuan Lyu","Minghui Liao","Cong Yao","Wenhao Wu","Xiang Bai"],"url":"https://arxiv.org/abs/1807.02242","pdfUrl":"https://arxiv.org/pdf/1807.02242","codeUrl":null,"preprint":{"id":"1807.02242","submittedOn":"2018-07-06","url":"https://arxiv.org/abs/1807.02242"},"publication":{"venue":"ECCV","year":2018,"kind":"conference","url":"https://arxiv.org/abs/1807.02242"},"citation":"@misc{textspottingmasktextspotter2018,\n  title = {{Mask TextSpotter: An End-to-End Trainable Neural Network for Spotting Text with Arbitrary Shapes}},\n  author = {Pengyuan Lyu and Minghui Liao and Cong Yao and Wenhao Wu and Xiang Bai},\n  year = {2018},\n  eprint = {1807.02242},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/1807.02242}\n}\n","citationPath":"/content/papers/citations/text-spotting/mask-textspotter.085f90e5d710.bib"},"thumbnail":{"src":"/content/papers/assets/c54febd257fcd2b211ed4cd3b062b2568f9feea65abeb983de0d5685dd36f2ce.jpg","width":1224,"height":448},"locale":"en","path":"/content/papers/en/text-spotting/mask-textspotter.d4735435d2a9.md","headings":[{"id":"problem-definition","text":"Problem Definition","depth":2},{"id":"solution","text":"Solution","depth":2},{"id":"model-architecture","text":"Model Architecture","depth":3},{"id":"backbone--neck","text":"Backbone & Neck","depth":3},{"id":"region-proposal-network-rpn","text":"Region Proposal Network (RPN)","depth":3},{"id":"fast-r-cnn-branch","text":"Fast R-CNN Branch","depth":3},{"id":"mask-branch","text":"Mask Branch","depth":3},{"id":"label-generation","text":"Label Generation","depth":3},{"id":"loss-function","text":"Loss Function","depth":3},{"id":"inference-phase","text":"Inference Phase","depth":3},{"id":"pixel-voting-algorithm","text":"Pixel Voting Algorithm","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"horizontal-text-recognition","text":"Horizontal Text Recognition","depth":3},{"id":"skewed-text-recognition","text":"Skewed Text Recognition","depth":3},{"id":"curved-text-recognition","text":"Curved Text Recognition","depth":3},{"id":"visualization","text":"Visualization","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"text-spotting/abcnet","relation":"next","topic":"Joint text detection and recognition"}]},{"id":"text-spotting/abcnet-v2","category":"text-spotting","legacyPath":"text-spotting/2105-abcnet-v2","legacySlug":"2105-abcnet-v2","legacyPaths":[],"title":"ABCNet v2","subtitle":"The One-bit Challenge","author":"Z. Yuan","paperDate":"2021-05-08","publishedOn":"2024-11-13","paperTitle":"ABCNet v2: Adaptive Bezier-Curve Network for Real-time End-to-end Text Spotting","paperUrl":"https://arxiv.org/abs/2105.03620","bibliography":{"title":"ABCNet v2: Adaptive Bezier-Curve Network for Real-time End-to-end Text Spotting","authors":["Yuliang Liu","Chunhua Shen","Lianwen Jin","Tong He","Peng Chen","Chongyu Liu","Hao Chen"],"url":"https://arxiv.org/abs/2105.03620","pdfUrl":"https://arxiv.org/pdf/2105.03620","codeUrl":null,"preprint":{"id":"2105.03620","submittedOn":"2021-05-08","url":"https://arxiv.org/abs/2105.03620"},"publication":{"venue":"IEEE Transactions on Pattern Analysis and Machine Intelligence","shortVenue":"IEEE TPAMI","year":2021,"kind":"journal","url":"https://doi.org/10.1109/tpami.2021.3107437"},"citation":"@misc{textspottingabcnetv22021,\n  title = {{ABCNet v2: Adaptive Bezier-Curve Network for Real-time End-to-end Text Spotting}},\n  author = {Yuliang Liu and Chunhua Shen and Lianwen Jin and Tong He and Peng Chen and Chongyu Liu and Hao Chen},\n  year = {2021},\n  eprint = {2105.03620},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2105.03620}\n}\n","citationPath":"/content/papers/citations/text-spotting/abcnet-v2.832a039d6996.bib"},"thumbnail":{"src":"/content/papers/assets/b2384e28cda3ab0da78f552abca6c544953e0992acc35b3ae911ce80fa810132.jpg","width":1728,"height":284},"locale":"en","path":"/content/papers/en/text-spotting/abcnet-v2.4075ae99a55f.md","headings":[{"id":"problem-definition","text":"Problem Definition","depth":2},{"id":"problem-solving-strategies","text":"Problem-Solving Strategies","depth":2},{"id":"model-architecture","text":"Model Architecture","depth":3},{"id":"coordinate-convolution-module","text":"Coordinate Convolution Module","depth":3},{"id":"attention-based-recognition-branch","text":"Attention-based Recognition Branch","depth":3},{"id":"adaptive-end-to-end-training-aet","text":"Adaptive End-to-End Training (AET)","depth":3},{"id":"model-quantization","text":"Model Quantization","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"ablation-study","text":"Ablation Study","depth":3},{"id":"comparison-of-text-detection-methods","text":"Comparison of Text Detection Methods","depth":3},{"id":"comparison-of-text-recognition-methods","text":"Comparison of Text Recognition Methods","depth":3},{"id":"quantized-model-performance","text":"Quantized Model Performance","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"text-spotting/abcnet","relation":"background","topic":"Joint text detection and recognition"}]}],"text":"OCR has been advancing for years, but text detection and recognition tasks remain challenging.\n\n## Defining the Problem\n\nText detection and recognition are the two primary tasks of OCR.\n\n- **Text Detection** identifies the location of text within an image. Many methods for this task are derived from object detection and segmentation techniques. However, text often appears distorted, slanted, or curved, making detection difficult.\n\n- **Text Recognition** converts detected text into recognizable characters. The performance of recognition is heavily influenced by detection errors. Therefore, many approaches assume irregular text as input, developing methods based on **STN (Spatial Transformer Networks)** or **2D attention mechanisms**.\n\nBelow is a recap of some past approaches:\n\n![ref](/content/papers/assets/15d28e59700249db752ab7fb4d6705c83fe9ba8c0449a125e6512ece70f7cefc.jpg)\n\nThe authors argue that complex multi-stage systems are slow and difficult to maintain in practice. To handle deformed text, they introduce **Bezier curves** to model its shape and **BezierAlign** to connect text detection with recognition.\n\nWait, what’s a **Bezier curve**?\n\n## Bezier Curves\n\n- [**Bézier curve - Wikipedia**](https://en.wikipedia.org/wiki/B%C3%A9zier_curve)\n\nA **Bezier curve** is a **mathematical curve** used to describe **smooth shapes**. It is widely applied in **computer graphics, typography, animation, and UI design** to create smooth, scalable curves. A Bezier curve is defined by a set of **control points** that determine how the curve bends. The first and last control points define the curve's endpoints, while the intermediate points shape the curve without necessarily lying on it.\n\nThese curves retain their smoothness under scaling, making them ideal for modeling various shapes, from straight lines to complex geometries.\n\nHere’s a glimpse of different types of Bezier curves:\n\n- **Linear Bezier Curve (Degree 1)**: Straight line between two control points.\n\n  - $B(t) = (1 - t) P_0 + t P_1, \\quad 0 \\leq t \\leq 1$\n\n    ![stage1](/content/papers/assets/d01918ffbf8d4e2a75578639385b4845384d8719471fe349395a54e710adc6c4.gif)\n\n- **Quadratic Bezier Curve (Degree 2)**: Uses three control points.\n\n  - $B(t) = (1 - t)^2 P_0 + 2(1 - t) t P_1 + t^2 P_2, \\quad 0 \\leq t \\leq 1$\n\n    ![stage2](/content/papers/assets/e445c56b1c3a3f6b5204147b1cf35d63a7c360f8819aa37a975e1ecf41370b2e.gif)\n\n- **Cubic Bezier Curve (Degree 3)**: Four control points, often used in fonts, animation, and UI.\n\n  - $B(t) = (1 - t)^3 P_0 + 3(1 - t)^2 t P_1 + 3(1 - t) t^2 P_2 + t^3 P_3$\n\n    ![stage3](/content/papers/assets/0e10d6da5983ab624704b394383ee3684093125e5dab5ebf4412fd6f0f939f78.gif)\n\n- **Quartic Bezier Curve (Degree 4)**: Five control points.\n\n  - $B(t) = (1 - t)^4 P_0 + 4(1 - t)^3 t P_1 + 6(1 - t)^2 t^2 P_2 + 4(1 - t) t^3 P_3 + t^4 P_4$\n\n    ![stage4](/content/papers/assets/67b905dac00c75bca37294098109756d9275465de4310a7a504e6b411b10e124.gif)\n\n- **Quintic Bezier Curve (Degree 5)**: Six control points.\n\n  - $B(t) = (1 - t)^5 P_0 + 5(1 - t)^4 t P_1 + 10(1 - t)^3 t^2 P_2 + 10(1 - t)^2 t^3 P_3 + 5(1 - t) t^4 P_4 + t^5 P_5$\n\n    ![stage5](/content/papers/assets/8fef2d6b16b7d1b6125040e11622352d0f3453e4b5960f092b4aea4092525654.gif)\n\n\n> [!NOTE]\n> \n> All these GIFs are from the Bezier curve section on Wikipedia.\n\n\n## Solution Overview\n\nAfter learning about Bezier curves, it seems like we’ve read the paper. (~Not! Yet!~)\n\n### Model Architecture\n\n![model architecture](/content/papers/assets/0d8681d7dcbb702c346101a5dc93b5f4ee4f05dcca5293b3ec7f5962c8cf6e0d.jpg)\n\nThe authors use a single-stage, **anchor-free** detector and Bezier curves to describe text shapes, forming an **end-to-end framework** for detecting and recognizing text of arbitrary shapes.\n\nThe authors find that cubic Bezier curves can represent most curved text shapes, so they use **8 control points** to represent the upper and lower boundaries of each text region.\n\n\n\n![bezier curve](/content/papers/assets/40c35d010cdd33742bf6583fe3721d02d937831d39ecc0fce70513f3228a7365.jpg)\n\n\n\nControl points are learned by **relative displacement** to simplify the model and improve accuracy:\n\n$$\n\\Delta x = b_i^x - x_{\\text{min}}, \\quad \\Delta y = b_i^y - y_{\\text{min}}\n$$\n\nHere, $x_{\\text{min}}$ and $y_{\\text{min}}$ are the minimum $x$ and $y$ values of the bounding box vertices.\n\nThe curve-control-point regression branch adds just **one convolutional layer** with 16 output channels to learn $\\Delta x$ and $\\Delta y$, adding little computation.\n\n\n> [!TIP]\n> \n> The text region is divided into upper and lower curves to enclose the text region. Therefore, a cubic Bezier curve requires 4 control points, and two curves together require 8 control points.\n> \n> Each point has coordinates $(x, y)$, so 8 control points have a total of 16 coordinates.\n\n\n### Generating GroundTruth\n\nSince existing datasets do not have Bezier curve annotations, the authors propose a method to generate **Bezier curve ground truth** from the polygonal annotations of Total-Text and CTW1500 using the **least squares method**.\n\nThe equation looks like this:\n\n$$\n\\begin{bmatrix}\nB_{0,3}(t_0) & \\cdots & B_{3,3}(t_0) \\\\\nB_{0,3}(t_1) & \\cdots & B_{3,3}(t_1) \\\\\n\\vdots & \\ddots & \\vdots \\\\\nB_{0,3}(t_m) & \\cdots & B_{3,3}(t_m)\n\\end{bmatrix}\n\\begin{bmatrix}\nb_{x0} & b_{y0} \\\\\nb_{x1} & b_{y1}\\\\\nb_{x2} & b_{y2}\\\\\nb_{x3} & b_{y3}\\\\\n\\end{bmatrix}\n=\n\\begin{bmatrix}\np_{x0} & p_{y0} \\\\\np_{x1} & p_{y1} \\\\\n\\vdots & \\vdots \\\\\np_{xm} & p_{ym}\n\\end{bmatrix}\n$$\n\nThe annotated points come from the upper and lower text boundaries, with a separate cubic Bezier curve fitted to each. The indices run from 0 through $m$, so there are $m+1$ annotated points in this expression. Each $t_i$ is determined by the cumulative length along its boundary polyline. An example follows:\n\n![bezier curve generation](/content/papers/assets/e8acb89398d3cb20e01a884f26956a5fb2ae40a717af0140537fad21cf65c240.jpg)\n\n### Synthetic Data\n\nTo train an end-to-end scene text detection model, large amounts of synthetic data can help. However, the existing **SynthText dataset** (containing 800,000 images) only provides quadrilateral bounding boxes for mostly straight text, which is insufficient for handling arbitrarily shaped scene text. To address this limitation, the authors expanded the dataset by synthesizing 150,000 diverse images to enhance the model's ability to detect curved text.\n\nThe composition of the new dataset is as follows:\n\n- **94,723 images**: Containing mostly **straight text**.\n- **54,327 images**: Containing mostly **curved text**.\n\nThe sources for these images are:\n\n- **VGG synthesis method**: Used to generate the synthetic images.\n- **40,000 background images**: Sourced from the **COCO-Text dataset**.\n\nEach background image from COCO-Text is prepared with a **segmentation mask**, along with **scene depth information**, to improve the blending of text with the background. To further increase the variety of text shapes, the authors modified the VGG synthesis method by incorporating **various artistic fonts** and a **rich text corpus**. They applied **curved distortions** and **diverse layouts** to the generated text, simulating a wide range of real-world text styles. For each text instance, **polygonal annotations** were automatically generated to describe the shape.\n\nAdditionally, the polygonal annotations were converted into **cubic Bezier curves** as ground truth, providing fitting targets for curved text shapes. These processes simplify the data generation workflow while effectively enhancing the model's ability to detect arbitrarily shaped text, meeting the demands of real-world applications.\n\nThe results of the synthetic data generation are shown below:\n\n![synthetic data](/content/papers/assets/5cf77d2b8d5a74812dba7c88e796aaa960c45d9e254bd13e6ca03b4148dba052.jpg)\n\n### Bezier Alignment\n\nTraditional sampling methods include:\n\n- [**RoI Pooling**](https://arxiv.org/abs/1707.03985)\n- [**RoI Rotate**](https://arxiv.org/abs/1801.01671)\n- [**Text-Align-Sampling**](https://arxiv.org/abs/1803.03474)\n- [**RoI Transform**](https://arxiv.org/abs/1812.09900)\n\n**BezierAlign** extends RoIAlign to curved regions. For each sampled curve parameter, it connects the corresponding top and bottom boundary points, then interpolates along that segment. The segments need not be perpendicular to the boundaries. Arc-length spacing also depends on the curve’s geometry. Bilinear interpolation reads features at the resulting positions.\n\nGiven a feature map and the control points of a Bezier curve, the pixel positions in the rectangular output feature map are calculated as follows:\n\nAssume the size of the output feature map is $h_{\\text{out}} \\times w_{\\text{out}}$.\n\nThe process is as follows:\n\n- For the $i$-th pixel $g_i$ its grid position is denoted as $(g_{iw}, g_{ih})$.\n- The parameter $t$ is computed using the formula:\n\n  $$\n  t = \\frac{g_{iw}}{w_{\\text{out}}}\n  $$\n\n- Using $t$ and the Bezier curve equation, the **top boundary point** $t_p$ and **bottom boundary point** $b_p$ are calculated.\n\n- Based on $t_p$ and $b_p$, the position of the sampling point $o_p$ is determined using:\n\n  $$\n  o_p = b_p \\cdot \\frac{g_{ih}}{h_{\\text{out}}} + t_p \\cdot \\left(1 - \\frac{g_{ih}}{h_{\\text{out}}}\\right)\n  $$\n\n- Once the position $o_p$ is obtained, **bilinear interpolation** is applied to compute the final value at that point.\n\n---\n\nBezierAlign samples features along curved text regions for the recognition branch. The figure below compares the positions of different sampling grids; the quantitative recognition scores appear later in the ablation study.\n\n![comparison](/content/papers/assets/636321131c3bd87826ae7298a063cda590e60b58491aedc31bc16219c0e0c761.jpg)\n\n### Text Recognition Head\n\n\n\n![head architecture](/content/papers/assets/ce415d0f9f3dd9c813e4757f62bcb90f682414bc11514026af8b28c9b0a46379.jpg)\n\n\n\nUsing the backbone features shared with the detection branch and BezierAlign, the authors build a **lightweight text recognition branch**, shown above:\n\n- **6 convolutional layers**: The first four use stride 1; the last two use vertical stride 2 and horizontal stride 1.\n- **Average pooling**: Compresses the height dimension to produce $(n, 256, 1, w)$.\n- **Channel permutation**: Reorders the tensor for the **BLSTM**.\n- **BLSTM**: Produces $(w, n, 512)$ and captures context in both sequence directions.\n- **Fully connected layer**: Outputs $(w, n, n_{\\text{class}})$, where $n_{\\text{class}} = 97$ covers uppercase and lowercase letters, digits, symbols, an other-symbol class, and the paper's special EOF class.\n\nDuring training, recognition features are sampled from ground-truth Bezier regions, so early detection-boundary errors do not directly affect recognition sampling. The branches still share a backbone and are trained jointly. **CTC loss** aligns recognition outputs with text annotations.\n\nDuring inference, the detected Bezier curves replace the ground-truth regions as the RoIs, aligning recognition features with the curved text.\n\n### Training Configuration\n\n- **Backbone**: ResNet-50 with a Feature Pyramid Network (FPN).\n- **RoIAlign scales reported in the paper’s implementation description**:\n  - Resolutions are $1/8$, $1/16$, $1/32$, $1/64$, and $1/128$ of the input image.\n- **BezierAlign**: Applied on feature maps of 3 different sizes:\n  - Sizes are $1/4$, $1/8$, and $1/16$ of the input image.\n- **Pretraining datasets**:\n  - 150,000 synthetic images.\n  - 15,000 selected images from the **COCO-Text dataset**.\n  - 7,000 images from the **ICDAR-MLT dataset**.\n\nThe model is **pretrained on public English word-level datasets** and fine-tuned on the target dataset.\n\n---\n\nData Augmentation Strategies\n\n- **Random scaling during training**:\n  - The shorter side of the input is randomly chosen between 560 and 800 pixels.\n  - The longer side is constrained to a maximum of 1333 pixels.\n- **Random cropping**:\n  - Ensures the cropped region covers at least half of the original image and does not truncate text.\n  - If the cropping conditions cannot be met, cropping is skipped.\n\n---\n\nHardware Setup\n\n- **4 Tesla V100 GPUs** are used for training.\n- **Batch size**: 32 images per batch.\n\n---\n\nTraining Parameters\n\n- **Maximum iterations**: 150,000.\n- **Initial learning rate**: 0.01.\n- **Learning rate schedule**:\n  - Reduced to **0.001** at 70,000 iterations.\n  - Further reduced to **0.0001** at 120,000 iterations.\n\nThe entire training process takes approximately **3 days** to complete.\n\n## Discussion\n\nThe authors conducted comparisons and ablation studies on the **Total-Text dataset** to evaluate the performance of their method. The results are as follows:\n\n### BezierAlign Analysis\n\n\n\n![bezier align](/content/papers/assets/73eec50d2c83f1dd9bbfc1a2b68e7a0fe1ad10c58bd5685e127ef9ec82cd5541.jpg)\n\n\n\nTo validate the effectiveness of **BezierAlign**, the authors performed ablation experiments on **Total-Text**, exploring the impact of different sampling point configurations, as shown in the table above.\n\nThe results show that the number of sampling points affects both performance and speed. With **(7, 32)** sampling points, the table reports **61.9% F-measure** at **22.8 FPS**, the accuracy–speed balance chosen by the authors for later experiments.\n\nCompared with earlier sampling methods, BezierAlign reaches **61.9% end-to-end text recognition F-measure**, versus **38.4%** for horizontal sampling and **44.7%** for quadrilateral sampling. Table 3 reports F-measure for this comparison, without an FPS measurement:\n\n\n\n![bezier align comparison](/content/papers/assets/14657fbe33c62ef40ea9d0721fd46563eef7f77dcb1101394b5c98e91055c915.jpg)\n\n\n\n### BezierCurve Analysis\n\n\n\n![bezier curve](/content/papers/assets/7b94569ab1fec1f0bc3a60f7a2c6d6630bbbbe8aacb2786742896eb800a3b442.jpg)\n\n\n\nWithout Bezier-curve detection, the table reports **22.8 FPS**; with it, **22.5 FPS**. The additional cost is small in this ablation measurement.\n\n### Experimental Results on Total-Text\n\n![total-text](/content/papers/assets/5d0bb6c7c2e15628ceef8afc5bfbc5abb814022ecd689b86fa1e47b860f318d5.jpg)\n\nThe authors compared **ABCNet** with other methods, as shown in the table above:\n\n- **Single-scale inference**: ABCNet scores **64.2% without a lexicon** and **75.7% with the full lexicon**, at **17.9 FPS**. The faster ABCNet-F scores **61.9%/74.1%** at **22.8 FPS**.\n- **Multi-scale inference**: ABCNet-MS reaches the table's highest scores, **69.5%/78.4%**, while speed falls to **6.9 FPS**. Training data and backbones differ across methods in this comparison.\n- **ABCNet-F versus Mask TextSpotter’19**: The speeds are **22.8 versus 2.0 FPS**, about **11.4 times** as fast, while the no-lexicon F-measures are **61.9% versus 65.3%**.\n\n## Conclusion\n\nABCNet introduces Bezier curves to an anchor-free detector and uses BezierAlign to align features with curved text boundaries for recognition. The authors also synthesize varied text images to help the model generalize.\n\nThe design philosophy of this paper provides new perspectives and directions for future text detection and recognition tasks, worthy of further research and exploration.\n","imageSizes":{"/content/papers/assets/15d28e59700249db752ab7fb4d6705c83fe9ba8c0449a125e6512ece70f7cefc.jpg":{"width":1708,"height":652},"/content/papers/assets/d01918ffbf8d4e2a75578639385b4845384d8719471fe349395a54e710adc6c4.gif":{"width":360,"height":150},"/content/papers/assets/e445c56b1c3a3f6b5204147b1cf35d63a7c360f8819aa37a975e1ecf41370b2e.gif":{"width":360,"height":150},"/content/papers/assets/0e10d6da5983ab624704b394383ee3684093125e5dab5ebf4412fd6f0f939f78.gif":{"width":360,"height":150},"/content/papers/assets/67b905dac00c75bca37294098109756d9275465de4310a7a504e6b411b10e124.gif":{"width":360,"height":150},"/content/papers/assets/8fef2d6b16b7d1b6125040e11622352d0f3453e4b5960f092b4aea4092525654.gif":{"width":473,"height":378},"/content/papers/assets/0d8681d7dcbb702c346101a5dc93b5f4ee4f05dcca5293b3ec7f5962c8cf6e0d.jpg":{"width":1940,"height":474},"/content/papers/assets/40c35d010cdd33742bf6583fe3721d02d937831d39ecc0fce70513f3228a7365.jpg":{"width":1224,"height":408},"/content/papers/assets/e8acb89398d3cb20e01a884f26956a5fb2ae40a717af0140537fad21cf65c240.jpg":{"width":1114,"height":554},"/content/papers/assets/5cf77d2b8d5a74812dba7c88e796aaa960c45d9e254bd13e6ca03b4148dba052.jpg":{"width":1224,"height":490},"/content/papers/assets/636321131c3bd87826ae7298a063cda590e60b58491aedc31bc16219c0e0c761.jpg":{"width":1988,"height":492},"/content/papers/assets/ce415d0f9f3dd9c813e4757f62bcb90f682414bc11514026af8b28c9b0a46379.jpg":{"width":880,"height":368},"/content/papers/assets/73eec50d2c83f1dd9bbfc1a2b68e7a0fe1ad10c58bd5685e127ef9ec82cd5541.jpg":{"width":844,"height":360},"/content/papers/assets/14657fbe33c62ef40ea9d0721fd46563eef7f77dcb1101394b5c98e91055c915.jpg":{"width":852,"height":196},"/content/papers/assets/7b94569ab1fec1f0bc3a60f7a2c6d6630bbbbe8aacb2786742896eb800a3b442.jpg":{"width":728,"height":156},"/content/papers/assets/5d0bb6c7c2e15628ceef8afc5bfbc5abb814022ecd689b86fa1e47b860f318d5.jpg":{"width":1224,"height":444}}}
