{
  "id": 413486,
  "title": "State-of-the-art methods: Curved text detection, Multi-oriented text detection and End-to-end text recognition.",
  "url": "/competitions/2023-kaggle-ai-report/discussion/413486",
  "author_name": "",
  "post_date": "2023-05-28T22:15:08.808221300Z",
  "votes": 20,
  "comment_count": 1,
  "views": 0,
  "content": "<h1>Curved text detection</h1>\n<p>CITATION<br>\n<a href=\"https://www.kaggle.com/unknown\" target=\"_blank\">@unknown</a>{unknown, author = {Sheng, Tao and Chen, Jie and Lian, Zhouhui},<br>\nyear = {2021}, month = {07}, pages = {}, <br>\ntitle = {CentripetalText: An Efficient Text Instance Representation for Scene Text Detection} }</p>\n<p>\"The authors first evaluated their Centripetal Text on the Total-Text and CTW1500 datasets to examine its capability for curved text detection. During testing, they set the short side of images to different scales (320, 512, 640) and kept their aspect ratios. They compared their methods with other state-of-the-art detectors.\"</p>\n<p>\"For Total-Text, when learning from scratch, CT-640 achieves the competitive F-measure of 84.9%, surpassing most existing methods pre-trained on external text datasets. When pre-training on SynthText, the F-measure value of their best model CT-640 reaches 86.3%, which is 0.6% better than second-best DRRG , while still ensuring the real-time detection speed (40.0 FPS).\"</p>\n<p>\"It was demonstrated that the accuracy-speed trade-off of some top-performing real-time text detectors, from which it can be observed that their CT (Centripetal Text) breaks through the limitation of accuracy-speed boundary. Analogous results can also be obtained on CTW1500. With external training data, the F-measure of CT-640 is 83.9%, the second place of all methods, which is only lower than DGGR.  Meanwhile, the speed can still exceed 40 FPS.\"</p>\n<p>\"In summary, the experiments conducted on Total-Text and CTW1500  demonstrated that the proposed CT (Centripetal Text) achieved superior or competitive results compared to state-of-the-art methods, indicating its superiority in modeling curved texts.\"</p>\n<h1>Multi-oriented text detection</h1>\n<p>\"The authors also evaluated CT (Centripetal Text) on the MSRA-TD500 dataset to test the robustness in modeling multi-oriented texts. CT achieved the F-measure value of 83.0% at 34.8 FPS without external training data. Compared with PAN, their method outperformed it by 4.1%. When pre-training on SynthText, the F-measure value of their CT can further be boosted to 86.1%. The highest performance and the fastest speed achieved by CT proved its generalization ability.\"</p>\n<h1>End-to-end text recognition</h1>\n<p>\"The authors simply replaced SPN ( Segmentation Proposal Network) in Mask TextSpotter v3 with our proposed CPN ( CentripetalText Proposal Network) to develop a more powerful end-to-end text recognizer. They evaluated CPN-based text spotter on Total-Text to test the proposal generation quality for the text spotting task. </p>\n<p>\"Mask TextSpotter v3 achieved the F-measure values of 71.9% and 79.5% when the lexicon is not used and used respectively. Compared with the original version and other state-of-the-art methods, their method can obtain higher performance whether the lexicon is provided or not. Thus, the quantitative results demonstrate that CPN can produce more accurate text proposals than<br>\nSPN, which is beneficial for recognition and can improve the performance of end-to-end text recognition further.\"</p>\n<p><a href=\"https://openreview.net/pdf?id=z1F9G4VnGZ-\" target=\"_blank\">https://openreview.net/pdf?id=z1F9G4VnGZ-</a></p>\n<h1>On Kaggle</h1>\n<p>Text recognition: Total Text dataset By Konrad Banachewicz<br>\n<a href=\"https://www.kaggle.com/datasets/konradb/text-recognition-total-text-dataset\" target=\"_blank\">https://www.kaggle.com/datasets/konradb/text-recognition-total-text-dataset</a></p>\n<h1>In fact, I didn't even understand this whole subject.</h1>\n<p>It's too complex for my poor knowledge about that stuff. Maybe it could be useful for those that are going to write any essay on this competition.</p>",
  "messages": [
    {
      "id": "2278692",
      "postDate": "05/28/2023 22:15:08",
      "content": "<h1>Curved text detection</h1>\n<p>CITATION<br>\n<a href=\"https://www.kaggle.com/unknown\" target=\"_blank\">@unknown</a>{unknown, author = {Sheng, Tao and Chen, Jie and Lian, Zhouhui},<br>\nyear = {2021}, month = {07}, pages = {}, <br>\ntitle = {CentripetalText: An Efficient Text Instance Representation for Scene Text Detection} }</p>\n<p>\"The authors first evaluated their Centripetal Text on the Total-Text and CTW1500 datasets to examine its capability for curved text detection. During testing, they set the short side of images to different scales (320, 512, 640) and kept their aspect ratios. They compared their methods with other state-of-the-art detectors.\"</p>\n<p>\"For Total-Text, when learning from scratch, CT-640 achieves the competitive F-measure of 84.9%, surpassing most existing methods pre-trained on external text datasets. When pre-training on SynthText, the F-measure value of their best model CT-640 reaches 86.3%, which is 0.6% better than second-best DRRG , while still ensuring the real-time detection speed (40.0 FPS).\"</p>\n<p>\"It was demonstrated that the accuracy-speed trade-off of some top-performing real-time text detectors, from which it can be observed that their CT (Centripetal Text) breaks through the limitation of accuracy-speed boundary. Analogous results can also be obtained on CTW1500. With external training data, the F-measure of CT-640 is 83.9%, the second place of all methods, which is only lower than DGGR.  Meanwhile, the speed can still exceed 40 FPS.\"</p>\n<p>\"In summary, the experiments conducted on Total-Text and CTW1500  demonstrated that the proposed CT (Centripetal Text) achieved superior or competitive results compared to state-of-the-art methods, indicating its superiority in modeling curved texts.\"</p>\n<h1>Multi-oriented text detection</h1>\n<p>\"The authors also evaluated CT (Centripetal Text) on the MSRA-TD500 dataset to test the robustness in modeling multi-oriented texts. CT achieved the F-measure value of 83.0% at 34.8 FPS without external training data. Compared with PAN, their method outperformed it by 4.1%. When pre-training on SynthText, the F-measure value of their CT can further be boosted to 86.1%. The highest performance and the fastest speed achieved by CT proved its generalization ability.\"</p>\n<h1>End-to-end text recognition</h1>\n<p>\"The authors simply replaced SPN ( Segmentation Proposal Network) in Mask TextSpotter v3 with our proposed CPN ( CentripetalText Proposal Network) to develop a more powerful end-to-end text recognizer. They evaluated CPN-based text spotter on Total-Text to test the proposal generation quality for the text spotting task. </p>\n<p>\"Mask TextSpotter v3 achieved the F-measure values of 71.9% and 79.5% when the lexicon is not used and used respectively. Compared with the original version and other state-of-the-art methods, their method can obtain higher performance whether the lexicon is provided or not. Thus, the quantitative results demonstrate that CPN can produce more accurate text proposals than<br>\nSPN, which is beneficial for recognition and can improve the performance of end-to-end text recognition further.\"</p>\n<p><a href=\"https://openreview.net/pdf?id=z1F9G4VnGZ-\" target=\"_blank\">https://openreview.net/pdf?id=z1F9G4VnGZ-</a></p>\n<h1>On Kaggle</h1>\n<p>Text recognition: Total Text dataset By Konrad Banachewicz<br>\n<a href=\"https://www.kaggle.com/datasets/konradb/text-recognition-total-text-dataset\" target=\"_blank\">https://www.kaggle.com/datasets/konradb/text-recognition-total-text-dataset</a></p>\n<h1>In fact, I didn't even understand this whole subject.</h1>\n<p>It's too complex for my poor knowledge about that stuff. Maybe it could be useful for those that are going to write any essay on this competition.</p>",
      "rawMarkdown": "#Curved text detection \n\n\nCITATION\n@unknown{unknown, author = {Sheng, Tao and Chen, Jie and Lian, Zhouhui},\nyear = {2021}, month = {07}, pages = {}, \ntitle = {CentripetalText: An Efficient Text Instance Representation for Scene Text Detection} }\n\n \"The authors first evaluated their Centripetal Text on the Total-Text and CTW1500 datasets to examine its capability for curved text detection. During testing, they set the short side of images to different scales (320, 512, 640) and kept their aspect ratios. They compared their methods with other state-of-the-art detectors.\"\n\n\"For Total-Text, when learning from scratch, CT-640 achieves the competitive F-measure of 84.9%, surpassing most existing methods pre-trained on external text datasets. When pre-training on SynthText, the F-measure value of their best model CT-640 reaches 86.3%, which is 0.6% better than second-best DRRG , while still ensuring the real-time detection speed (40.0 FPS).\"\n\n\"It was demonstrated that the accuracy-speed trade-off of some top-performing real-time text detectors, from which it can be observed that their CT (Centripetal Text) breaks through the limitation of accuracy-speed boundary. Analogous results can also be obtained on CTW1500. With external training data, the F-measure of CT-640 is 83.9%, the second place of all methods, which is only lower than DGGR.  Meanwhile, the speed can still exceed 40 FPS.\"\n\n\"In summary, the experiments conducted on Total-Text and CTW1500  demonstrated that the proposed CT (Centripetal Text) achieved superior or competitive results compared to state-of-the-art methods, indicating its superiority in modeling curved texts.\"\n\n#Multi-oriented text detection\n\n\"The authors also evaluated CT (Centripetal Text) on the MSRA-TD500 dataset to test the robustness in modeling multi-oriented texts. CT achieved the F-measure value of 83.0% at 34.8 FPS without external training data. Compared with PAN, their method outperformed it by 4.1%. When pre-training on SynthText, the F-measure value of their CT can further be boosted to 86.1%. The highest performance and the fastest speed achieved by CT proved its generalization ability.\"\n\n#End-to-end text recognition\n\n\"The authors simply replaced SPN ( Segmentation Proposal Network) in Mask TextSpotter v3 with our proposed CPN ( CentripetalText Proposal Network) to develop a more powerful end-to-end text recognizer. They evaluated CPN-based text spotter on Total-Text to test the proposal generation quality for the text spotting task. \n\n\"Mask TextSpotter v3 achieved the F-measure values of 71.9% and 79.5% when the lexicon is not used and used respectively. Compared with the original version and other state-of-the-art methods, their method can obtain higher performance whether the lexicon is provided or not. Thus, the quantitative results demonstrate that CPN can produce more accurate text proposals than\nSPN, which is beneficial for recognition and can improve the performance of end-to-end text recognition further.\"\n\nhttps://openreview.net/pdf?id=z1F9G4VnGZ-\n\n#On Kaggle\n\nText recognition: Total Text dataset By Konrad Banachewicz\nhttps://www.kaggle.com/datasets/konradb/text-recognition-total-text-dataset\n\n#In fact, I didn't even understand this whole subject. \n\nIt's too complex for my poor knowledge about that stuff. Maybe it could be useful for those that are going to write any essay on this competition.",
      "votes": null
    },
    {
      "id": "2294292",
      "postDate": "06/10/2023 00:10:22",
      "content": "<p>This article highlights three state-of-the-art methods in the field of text detection and recognition: curved text detection, multi-oriented text detection, and end-to-end text recognition. These methods have been evaluated and compared with existing approaches to showcase their capabilities and advantages.</p>\n<p>For curved text detection, the authors introduced CentripetalText, which demonstrated excellent performance on the Total-Text and CTW1500 datasets. The results showed that CentripetalText achieved competitive F-measure values, surpassing most existing methods, even when learning from scratch or pre-training on SynthText. The proposed CentripetalText method broke through the accuracy-speed boundary, providing superior modeling of curved texts.</p>\n<p>In multi-oriented text detection, CentripetalText was also evaluated on the MSRA-TD500 dataset. The results exhibited the robustness of CentripetalText in modeling multi-oriented texts, outperforming the PAN method by 4.1%. With or without external training data, CentripetalText achieved high F-measure values and demonstrated excellent generalization ability.</p>\n<p>Regarding end-to-end text recognition, the authors replaced the Segmentation Proposal Network (SPN) with the proposed CentripetalText Proposal Network (CPN) in Mask TextSpotter v3. The CPN-based text spotter showcased improved proposal generation quality, resulting in higher performance compared to the original version and other state-of-the-art methods. This enhancement in text proposal accuracy was found to be beneficial for recognition and further improved the performance of end-to-end text recognition.</p>\n<p>While the subject matter may seem complex, these advancements in text detection and recognition have significant implications for various applications, including document processing, optical character recognition (OCR), and text-based data analysis. Researchers and practitioners in the field can utilize these state-of-the-art methods to enhance their understanding and contribute to the development of more accurate and efficient text detection and recognition systems.</p>\n<p>For those interested in delving deeper into this topic or considering writing an essay or participating in a competition, exploring the provided citations and datasets can serve as valuable resources. It's important to acknowledge the complexity of the subject matter and the ongoing advancements in the field, but with curiosity and dedication, one can gradually gain a deeper understanding of these technologies and their potential applications.</p>",
      "rawMarkdown": "This article highlights three state-of-the-art methods in the field of text detection and recognition: curved text detection, multi-oriented text detection, and end-to-end text recognition. These methods have been evaluated and compared with existing approaches to showcase their capabilities and advantages.\n\nFor curved text detection, the authors introduced CentripetalText, which demonstrated excellent performance on the Total-Text and CTW1500 datasets. The results showed that CentripetalText achieved competitive F-measure values, surpassing most existing methods, even when learning from scratch or pre-training on SynthText. The proposed CentripetalText method broke through the accuracy-speed boundary, providing superior modeling of curved texts.\n\nIn multi-oriented text detection, CentripetalText was also evaluated on the MSRA-TD500 dataset. The results exhibited the robustness of CentripetalText in modeling multi-oriented texts, outperforming the PAN method by 4.1%. With or without external training data, CentripetalText achieved high F-measure values and demonstrated excellent generalization ability.\n\nRegarding end-to-end text recognition, the authors replaced the Segmentation Proposal Network (SPN) with the proposed CentripetalText Proposal Network (CPN) in Mask TextSpotter v3. The CPN-based text spotter showcased improved proposal generation quality, resulting in higher performance compared to the original version and other state-of-the-art methods. This enhancement in text proposal accuracy was found to be beneficial for recognition and further improved the performance of end-to-end text recognition.\n\nWhile the subject matter may seem complex, these advancements in text detection and recognition have significant implications for various applications, including document processing, optical character recognition (OCR), and text-based data analysis. Researchers and practitioners in the field can utilize these state-of-the-art methods to enhance their understanding and contribute to the development of more accurate and efficient text detection and recognition systems.\n\nFor those interested in delving deeper into this topic or considering writing an essay or participating in a competition, exploring the provided citations and datasets can serve as valuable resources. It's important to acknowledge the complexity of the subject matter and the ongoing advancements in the field, but with curiosity and dedication, one can gradually gain a deeper understanding of these technologies and their potential applications.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2294292,
      "author_name": "ahmadihossein",
      "author_url": "",
      "post_date": "06/10/2023 00:10:22",
      "content": "<p>This article highlights three state-of-the-art methods in the field of text detection and recognition: curved text detection, multi-oriented text detection, and end-to-end text recognition. These methods have been evaluated and compared with existing approaches to showcase their capabilities and advantages.</p>\n<p>For curved text detection, the authors introduced CentripetalText, which demonstrated excellent performance on the Total-Text and CTW1500 datasets. The results showed that CentripetalText achieved competitive F-measure values, surpassing most existing methods, even when learning from scratch or pre-training on SynthText. The proposed CentripetalText method broke through the accuracy-speed boundary, providing superior modeling of curved texts.</p>\n<p>In multi-oriented text detection, CentripetalText was also evaluated on the MSRA-TD500 dataset. The results exhibited the robustness of CentripetalText in modeling multi-oriented texts, outperforming the PAN method by 4.1%. With or without external training data, CentripetalText achieved high F-measure values and demonstrated excellent generalization ability.</p>\n<p>Regarding end-to-end text recognition, the authors replaced the Segmentation Proposal Network (SPN) with the proposed CentripetalText Proposal Network (CPN) in Mask TextSpotter v3. The CPN-based text spotter showcased improved proposal generation quality, resulting in higher performance compared to the original version and other state-of-the-art methods. This enhancement in text proposal accuracy was found to be beneficial for recognition and further improved the performance of end-to-end text recognition.</p>\n<p>While the subject matter may seem complex, these advancements in text detection and recognition have significant implications for various applications, including document processing, optical character recognition (OCR), and text-based data analysis. Researchers and practitioners in the field can utilize these state-of-the-art methods to enhance their understanding and contribute to the development of more accurate and efficient text detection and recognition systems.</p>\n<p>For those interested in delving deeper into this topic or considering writing an essay or participating in a competition, exploring the provided citations and datasets can serve as valuable resources. It's important to acknowledge the complexity of the subject matter and the ongoing advancements in the field, but with curiosity and dedication, one can gradually gain a deeper understanding of these technologies and their potential applications.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2278692": "#Curved text detection \n\n\nCITATION\n@unknown{unknown, author = {Sheng, Tao and Chen, Jie and Lian, Zhouhui},\nyear = {2021}, month = {07}, pages = {}, \ntitle = {CentripetalText: An Efficient Text Instance Representation for Scene Text Detection} }\n\n \"The authors first evaluated their Centripetal Text on the Total-Text and CTW1500 datasets to examine its capability for curved text detection. During testing, they set the short side of images to different scales (320, 512, 640) and kept their aspect ratios. They compared their methods with other state-of-the-art detectors.\"\n\n\"For Total-Text, when learning from scratch, CT-640 achieves the competitive F-measure of 84.9%, surpassing most existing methods pre-trained on external text datasets. When pre-training on SynthText, the F-measure value of their best model CT-640 reaches 86.3%, which is 0.6% better than second-best DRRG , while still ensuring the real-time detection speed (40.0 FPS).\"\n\n\"It was demonstrated that the accuracy-speed trade-off of some top-performing real-time text detectors, from which it can be observed that their CT (Centripetal Text) breaks through the limitation of accuracy-speed boundary. Analogous results can also be obtained on CTW1500. With external training data, the F-measure of CT-640 is 83.9%, the second place of all methods, which is only lower than DGGR.  Meanwhile, the speed can still exceed 40 FPS.\"\n\n\"In summary, the experiments conducted on Total-Text and CTW1500  demonstrated that the proposed CT (Centripetal Text) achieved superior or competitive results compared to state-of-the-art methods, indicating its superiority in modeling curved texts.\"\n\n#Multi-oriented text detection\n\n\"The authors also evaluated CT (Centripetal Text) on the MSRA-TD500 dataset to test the robustness in modeling multi-oriented texts. CT achieved the F-measure value of 83.0% at 34.8 FPS without external training data. Compared with PAN, their method outperformed it by 4.1%. When pre-training on SynthText, the F-measure value of their CT can further be boosted to 86.1%. The highest performance and the fastest speed achieved by CT proved its generalization ability.\"\n\n#End-to-end text recognition\n\n\"The authors simply replaced SPN ( Segmentation Proposal Network) in Mask TextSpotter v3 with our proposed CPN ( CentripetalText Proposal Network) to develop a more powerful end-to-end text recognizer. They evaluated CPN-based text spotter on Total-Text to test the proposal generation quality for the text spotting task. \n\n\"Mask TextSpotter v3 achieved the F-measure values of 71.9% and 79.5% when the lexicon is not used and used respectively. Compared with the original version and other state-of-the-art methods, their method can obtain higher performance whether the lexicon is provided or not. Thus, the quantitative results demonstrate that CPN can produce more accurate text proposals than\nSPN, which is beneficial for recognition and can improve the performance of end-to-end text recognition further.\"\n\nhttps://openreview.net/pdf?id=z1F9G4VnGZ-\n\n#On Kaggle\n\nText recognition: Total Text dataset By Konrad Banachewicz\nhttps://www.kaggle.com/datasets/konradb/text-recognition-total-text-dataset\n\n#In fact, I didn't even understand this whole subject. \n\nIt's too complex for my poor knowledge about that stuff. Maybe it could be useful for those that are going to write any essay on this competition.",
    "2294292": "This article highlights three state-of-the-art methods in the field of text detection and recognition: curved text detection, multi-oriented text detection, and end-to-end text recognition. These methods have been evaluated and compared with existing approaches to showcase their capabilities and advantages.\n\nFor curved text detection, the authors introduced CentripetalText, which demonstrated excellent performance on the Total-Text and CTW1500 datasets. The results showed that CentripetalText achieved competitive F-measure values, surpassing most existing methods, even when learning from scratch or pre-training on SynthText. The proposed CentripetalText method broke through the accuracy-speed boundary, providing superior modeling of curved texts.\n\nIn multi-oriented text detection, CentripetalText was also evaluated on the MSRA-TD500 dataset. The results exhibited the robustness of CentripetalText in modeling multi-oriented texts, outperforming the PAN method by 4.1%. With or without external training data, CentripetalText achieved high F-measure values and demonstrated excellent generalization ability.\n\nRegarding end-to-end text recognition, the authors replaced the Segmentation Proposal Network (SPN) with the proposed CentripetalText Proposal Network (CPN) in Mask TextSpotter v3. The CPN-based text spotter showcased improved proposal generation quality, resulting in higher performance compared to the original version and other state-of-the-art methods. This enhancement in text proposal accuracy was found to be beneficial for recognition and further improved the performance of end-to-end text recognition.\n\nWhile the subject matter may seem complex, these advancements in text detection and recognition have significant implications for various applications, including document processing, optical character recognition (OCR), and text-based data analysis. Researchers and practitioners in the field can utilize these state-of-the-art methods to enhance their understanding and contribute to the development of more accurate and efficient text detection and recognition systems.\n\nFor those interested in delving deeper into this topic or considering writing an essay or participating in a competition, exploring the provided citations and datasets can serve as valuable resources. It's important to acknowledge the complexity of the subject matter and the ongoing advancements in the field, but with curiosity and dedication, one can gradually gain a deeper understanding of these technologies and their potential applications."
  },
  "source": "meta"
}