Hi everyone,
\nThank you for all the valuable information about summarizing the previous threads and listing potential relevant literature to this competition.
\nBuilding upon @sytuannguyen's discussion thread about some relevant literature, I would like to start this discussion thread to summarize all the previous literature on plant specimen classification. I hope this could serve as a reference point for you guys when building models for our competition.
\nFocusing on the following questions:
\nTo my knowledge, state-of-the-art methods are most likely from our previous competition (Please correct me if I am wrong), e.g., herbarium 2020 and herbarium 2021. Therefore, for those of you who are running out of time, please refer to the paper about herbarium 2021. It contains a detailed description of models, losses, and schedulers to achieve state-of-the-art performance for herbarium specimen classification. Unfortunately, we don't have a publication for herbarium 2020, but many winning teams posted their solutions on the previous winning solutions. Many thanks to @init27 for the summary.
\nGoing deeper in the automated identification of herbarium specimens (Carranza-Rojas et al., 2017)
\n255 taxa totaling 11K images. Top-1 accruacy: 0.745 Top-5 accuracy: 0.872
\n1204 taxa totaling 292K images for training and 51K images for testing. Top-1 accuracy: 0.796, Top-5 accuracy:0.903
\nAutomated plant species identification - Trends and future directions (Waldchen et al., 2018)
Please feel free to extend this thread as you read other papers about plant specimen identification!
", "messages": [ { "id": 1697792, "postDate": "2022-02-19T20:57:43.973Z", "content": "Hi everyone,
\nThank you for all the valuable information about summarizing the previous threads and listing potential relevant literature to this competition.
\nBuilding upon @sytuannguyen's discussion thread about some relevant literature, I would like to start this discussion thread to summarize all the previous literature on plant specimen classification. I hope this could serve as a reference point for you guys when building models for our competition.
\nFocusing on the following questions:
\nTo my knowledge, state-of-the-art methods are most likely from our previous competition (Please correct me if I am wrong), e.g., herbarium 2020 and herbarium 2021. Therefore, for those of you who are running out of time, please refer to the paper about herbarium 2021. It contains a detailed description of models, losses, and schedulers to achieve state-of-the-art performance for herbarium specimen classification. Unfortunately, we don't have a publication for herbarium 2020, but many winning teams posted their solutions on the previous winning solutions. Many thanks to @init27 for the summary.
\nGoing deeper in the automated identification of herbarium specimens (Carranza-Rojas et al., 2017)
\n255 taxa totaling 11K images. Top-1 accruacy: 0.745 Top-5 accuracy: 0.872
\n1204 taxa totaling 292K images for training and 51K images for testing. Top-1 accuracy: 0.796, Top-5 accuracy:0.903
\nAutomated plant species identification - Trends and future directions (Waldchen et al., 2018)
Please feel free to extend this thread as you read other papers about plant specimen identification!
", "rawMarkdown": "Hi everyone,\n\nThank you for all the valuable information about summarizing the previous threads and listing potential relevant literature to this competition.\n\nBuilding upon @sytuannguyen's discussion thread about [some relevant literature](https://www.kaggle.com/c/herbarium-2022-fgvc9/discussion/307810), I would like to start this discussion thread to summarize all the previous literature on plant specimen classification. I hope this could serve as a reference point for you guys when building models for our competition. \n\nFocusing on the following questions:\n- What dataset do they use: How many taxa to classify? How many training images and testing images?\n- What are the models and the settings (backbone, loss, scheduler, etc.)?\n- What is the accuracy of their model?\n\nTo my knowledge, state-of-the-art methods are most likely from our previous competition (Please correct me if I am wrong), e.g., [herbarium 2020](https://www.kaggle.com/c/herbarium-2020-fgvc7) and [herbarium 2021](https://www.kaggle.com/c/herbarium-2021-fgvc8). Therefore, for those of you who are running out of time, please refer to the paper about [herbarium 2021](https://www.frontiersin.org/articles/10.3389/fpls.2021.787127/full). It contains a detailed description of models, losses, and schedulers to achieve state-of-the-art performance for herbarium specimen classification. Unfortunately, we don't have a publication for herbarium 2020, but many winning teams posted their solutions on the [previous winning solutions](https://www.kaggle.com/c/herbarium-2022-fgvc9/discussion/307624). Many thanks to @init27 for the summary.\n\n\n### Summary of the previous plant specimen classification results. \n\n[Computer vision applied to herbarium specimens of German trees: testing the future utility of the millions of herbarium specimen images for automated identification (Unger et al., 2016)](https://bmcecolevol.biomedcentral.com/articles/10.1186/s12862-016-0827-5)\n- 26 most common taxa in Germany, totaling 260 images for training. 0.72- 0.85 Top-1 accuracy\n - SVM with Fourier descriptors, multiple leaf shape parameters, weighted orientation histogram.\n\n[Going deeper in the automated identification of herbarium specimens (Carranza-Rojas et al., 2017)](https://bmcecolevol.biomedcentral.com/articles/10.1186/s12862-017-1014-z)\n- 255 taxa totaling 11K images. Top-1 accruacy: 0.745 Top-5 accuracy: 0.872\n - input resolution: 256 by 256. \n - best model pre-trained on imageNet and Herbarium1K dataset\n - GoogleNet model that stacks inception layers\n - Batch Normalization\n - PReLU\n - simple crop and resize data augmentation, default with Caffe.\n \n- 1204 taxa totaling 292K images for training and 51K images for testing. Top-1 accuracy: 0.796, Top-5 accuracy:0.903\n - Same setting\n\n[Automated plant species identification - Trends and future directions (Waldchen et al., 2018)](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1005993)