Congratulations to all the winners! We would like to thank the organizers for hosting such an engaging competition. While our team was able to achieve higher correlation results in plant traits, we unfortunately did not have a direct solution to obtain accurate values of these traits through plant images alone. By closely examining the images, we found that some plant traits are closely related to their growth period, making it challenging to obtain precise values using only mean and variance data. Additionally, there may be variations in scales or data units within the given data, making it difficult to model accurately. We hope that this competition can be held regularly in the future.
\nBelow is a brief explanation of our program:
\nWe attempted to conduct further regression analysis on plant traits after classification, but due to the limited number of samples per category, our results were unsatisfactory.
\nBelow is a link to our main codebase in the GitHub repository for our proposal on this challenge.
\nhttps://github.com/DuanChenL/FGVC10.
Fine-grained visual analysis is of immense practical importance, and while our current results are largely dependent on the accuracy of classification, we hope to continue exploring new methods for capturing plant trait information through images. Thank you again to the competition organizers for enriching our intellectual perspectives.
\n\n", "messages": [ { "id": 2312596, "postDate": "2023-06-22T03:54:53.620Z", "content": "Reference
\n[1] Woo S, Debnath S, Hu R, et al. ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders[J]. arXiv preprint arXiv:2301.00808, 2023.
\n
\n [2] S. Yun, D. Han, S. Chun, S. J. Oh, Y. Yoo and J. Choe, \"CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features,\" 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea (South), 2019, pp. 6022-6031, doi: 10.1109/ICCV.2019.00612.
Congratulations to all the winners! We would like to thank the organizers for hosting such an engaging competition. While our team was able to achieve higher correlation results in plant traits, we unfortunately did not have a direct solution to obtain accurate values of these traits through plant images alone. By closely examining the images, we found that some plant traits are closely related to their growth period, making it challenging to obtain precise values using only mean and variance data. Additionally, there may be variations in scales or data units within the given data, making it difficult to model accurately. We hope that this competition can be held regularly in the future.
\nBelow is a brief explanation of our program:
\nWe attempted to conduct further regression analysis on plant traits after classification, but due to the limited number of samples per category, our results were unsatisfactory.
\nBelow is a link to our main codebase in the GitHub repository for our proposal on this challenge.
\nhttps://github.com/DuanChenL/FGVC10.
Fine-grained visual analysis is of immense practical importance, and while our current results are largely dependent on the accuracy of classification, we hope to continue exploring new methods for capturing plant trait information through images. Thank you again to the competition organizers for enriching our intellectual perspectives.
\n\n", "rawMarkdown": "# **1st place solution**\nCongratulations to all the winners! We would like to thank the organizers for hosting such an engaging competition. While our team was able to achieve higher correlation results in plant traits, we unfortunately did not have a direct solution to obtain accurate values of these traits through plant images alone. By closely examining the images, we found that some plant traits are closely related to their growth period, making it challenging to obtain precise values using only mean and variance data. Additionally, there may be variations in scales or data units within the given data, making it difficult to model accurately. We hope that this competition can be held regularly in the future.\n### **Summary**\nBelow is a brief explanation of our program:\n\n1. We leveraged OpenCV to read and write corrupt images to our file system, thereby resolving the issue.\n2. Given the small sample size of each category, our data augmentation techniques included RandomResizedCrop, Transpose, HorizontalFlip, VerticalFlip, PiecewiseAffine, HueSaturationValue, RandomBrightnessContrast, and CutMix[2].\n3. We utilized the convnextv2_large[1] model from the timm model library as our backbone network for robust classification.\n4. Mean trait values for different plant categories were obtained using the data_train_mean.csv file and stored in label2num.json.\n5. We concatenated image features with meta data and utilized an MLP to obtain the final result.\n6. To obtain multiple models, we varied probabilities for data addition and model parameter dropout before ensembling them for the final classification results.\n7. Finally, we used label2num.json to query plant traits and obtain the final result.\n### **Failed attempt**\nWe attempted to conduct further regression analysis on plant traits after classification, but due to the limited number of samples per category, our results were unsatisfactory.\n###**Code**\nBelow is a link to our main codebase in the GitHub repository for our proposal on this challenge.\nhttps://github.com/DuanChenL/FGVC10.\n### **Conclusion**\nFine-grained visual analysis is of immense practical importance, and while our current results are largely dependent on the accuracy of classification, we hope to continue exploring new methods for capturing plant trait information through images. Thank you again to the competition organizers for enriching our intellectual perspectives. \n>### **Reference**\n[1] Woo S, Debnath S, Hu R, et al. ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders[J]. arXiv preprint arXiv:2301.00808, 2023.\n[2] S. Yun, D. Han, S. Chun, S. J. Oh, Y. Yoo and J. Choe, \"CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features,\" 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea (South), 2019, pp. 6022-6031, doi: 10.1109/ICCV.2019.00612.\n\n", "votes": 3 } ], "comments": [], "raw_markdown_by_id": { "2312596": "# **1st place solution**\nCongratulations to all the winners! We would like to thank the organizers for hosting such an engaging competition. While our team was able to achieve higher correlation results in plant traits, we unfortunately did not have a direct solution to obtain accurate values of these traits through plant images alone. By closely examining the images, we found that some plant traits are closely related to their growth period, making it challenging to obtain precise values using only mean and variance data. Additionally, there may be variations in scales or data units within the given data, making it difficult to model accurately. We hope that this competition can be held regularly in the future.\n### **Summary**\nBelow is a brief explanation of our program:\n\n1. We leveraged OpenCV to read and write corrupt images to our file system, thereby resolving the issue.\n2. Given the small sample size of each category, our data augmentation techniques included RandomResizedCrop, Transpose, HorizontalFlip, VerticalFlip, PiecewiseAffine, HueSaturationValue, RandomBrightnessContrast, and CutMix[2].\n3. We utilized the convnextv2_large[1] model from the timm model library as our backbone network for robust classification.\n4. Mean trait values for different plant categories were obtained using the data_train_mean.csv file and stored in label2num.json.\n5. We concatenated image features with meta data and utilized an MLP to obtain the final result.\n6. To obtain multiple models, we varied probabilities for data addition and model parameter dropout before ensembling them for the final classification results.\n7. Finally, we used label2num.json to query plant traits and obtain the final result.\n### **Failed attempt**\nWe attempted to conduct further regression analysis on plant traits after classification, but due to the limited number of samples per category, our results were unsatisfactory.\n###**Code**\nBelow is a link to our main codebase in the GitHub repository for our proposal on this challenge.\nhttps://github.com/DuanChenL/FGVC10.\n### **Conclusion**\nFine-grained visual analysis is of immense practical importance, and while our current results are largely dependent on the accuracy of classification, we hope to continue exploring new methods for capturing plant trait information through images. Thank you again to the competition organizers for enriching our intellectual perspectives. \n>### **Reference**\n[1] Woo S, Debnath S, Hu R, et al. ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders[J]. arXiv preprint arXiv:2301.00808, 2023.\n[2] S. Yun, D. Han, S. Chun, S. J. Oh, Y. Yoo and J. Choe, \"CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features,\" 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea (South), 2019, pp. 6022-6031, doi: 10.1109/ICCV.2019.00612.\n\n" } }Reference
\n[1] Woo S, Debnath S, Hu R, et al. ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders[J]. arXiv preprint arXiv:2301.00808, 2023.
\n
\n [2] S. Yun, D. Han, S. Chun, S. J. Oh, Y. Yoo and J. Choe, \"CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features,\" 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea (South), 2019, pp. 6022-6031, doi: 10.1109/ICCV.2019.00612.