Hello guys!
\nI noticed that many training image has a clutterred background(The training image is collected from a wild environment), and the surrounding plants in some images are not Cassava Leaf. Based on such observation, i reckon that the contextual information may lead to a biased model, especially when we use random crop to augment trianing image. In this setting, croped image may lost important discriminative feature of Cassava leaf but fill with many noised contextual information.
Hello guys!
\nI noticed that many training image has a clutterred background(The training image is collected from a wild environment), and the surrounding plants in some images are not Cassava Leaf. Based on such observation, i reckon that the contextual information may lead to a biased model, especially when we use random crop to augment trianing image. In this setting, croped image may lost important discriminative feature of Cassava leaf but fill with many noised contextual information.
Therefore, attention mode is needed. A model with attention function will help to better identify the characteristics of cassava and then classify
", "rawMarkdown": "Therefore, **attention** mode is needed. A model with **attention** function will help to better identify the characteristics of cassava and then classify", "votes": null }, { "id": "1149752", "postDate": "01/12/2021 05:41:21", "content": "totally, agree
", "rawMarkdown": "totally, agree", "votes": null }, { "id": "1150154", "postDate": "01/12/2021 12:15:15", "content": "One could segment out just the cassava leaves, which might help. I guess it's quite tedious to do, because someone has to annotate training data, which is a pretty unpleasant task.
\nAdditionally, we know that some of the images don't actually show images, but roots. At least those images are relatively easy to separate out from the ones that primarily show leaves, so it should be possible to treat those differently.
\nIn principle though, this seems like a segmentation taks that something like U-Net should do a prety good job at.
", "rawMarkdown": "One could segment out just the cassava leaves, which might help. I guess it's quite tedious to do, because someone has to annotate training data, which is a pretty unpleasant task.\n\nAdditionally, [we know](https://www.kaggle.com/bjoernholzhauer/cassava-leaf-disease-classif-eda-cv-strategy) that some of the images don't actually show images, but roots. At least those images are relatively easy to separate out from the ones that primarily show leaves, so it should be possible to treat those differently.\n\nIn principle though, this seems like a segmentation taks that something like U-Net should do a prety good job at.", "votes": null }, { "id": "1150954", "postDate": "01/13/2021 02:07:11", "content": "Hello, Björn. Thanks for your opinion. Segmenting images is tedious in this setting , like you said. what's more, backgound information is useful for image classification (NOISE OR SIGNAL: THE ROLE OF IMAGE BACKGROUNDS IN OBJECT RECOGNITION). I was thinking that to help model focus on saliency object, less consider the backgounds.
", "rawMarkdown": "Hello, Björn. Thanks for your opinion. Segmenting images is tedious in this setting , like you said. what's more, backgound information is useful for image classification ([NOISE OR SIGNAL: THE ROLE OF IMAGE BACKGROUNDS IN OBJECT RECOGNITION](https://openreview.net/pdf?id=gl3D-xY7wLq)). I was thinking that to help model focus on saliency object, less consider the backgounds.", "votes": null }, { "id": "1151132", "postDate": "01/13/2021 06:44:24", "content": "I think the background is quite noisy. We can use instance segmentation to segment the important areas, but we need unsupervised training. Of course, we don't need much precision. We roughly segment the important feature areas, and then add the attention element to identify the specific categories more accurately. But we also need to do this synchronously when predicting. So there are at least three problems that are not easy to solve
\nThe program is too time-consuming, and it may not be more accurate by simple classification + fusion?
Label problem cannot be solved, so unsupervised learning is needed
Unsupervised instance segmentation. I don't know much about this field. I don't know if there is a suitable network?
Another point is that even if you are more accurate than the final test set, there are some inaccurate labels on the test set, which also leads to some differences in the final score
", "rawMarkdown": "I think the background is quite noisy. We can use instance segmentation to segment the important areas, but we need unsupervised training. Of course, we don't need much precision. We roughly segment the important feature areas, and then add the attention element to identify the specific categories more accurately. But we also need to do this synchronously when predicting. So there are at least three problems that are not easy to solve\n\n1. The program is too time-consuming, and it may not be more accurate by simple classification + fusion?\n\n2. Label problem cannot be solved, so unsupervised learning is needed\n\n3. Unsupervised instance segmentation. I don't know much about this field. I don't know if there is a suitable network?\n\nAnother point is that even if you are more accurate than the final test set, there are some inaccurate labels on the test set, which also leads to some differences in the final score", "votes": null }, { "id": "1151622", "postDate": "01/13/2021 13:08:11", "content": "A simple center crop with a reasonable cropping size might also work.
", "rawMarkdown": "A simple center crop with a reasonable cropping size might also work.", "votes": null } ], "comments": [ { "id": 1149735, "author_name": "zhangeng", "author_url": "", "post_date": "01/12/2021 05:13:16", "content": "Therefore, attention mode is needed. A model with attention function will help to better identify the characteristics of cassava and then classify
", "votes": null, "replies": [ { "id": 1149752, "author_name": "wantsu", "author_url": "", "post_date": "01/12/2021 05:41:21", "content": "totally, agree
", "votes": null, "replies": [] } ] }, { "id": 1150154, "author_name": "bjoernholzhauer", "author_url": "", "post_date": "01/12/2021 12:15:15", "content": "One could segment out just the cassava leaves, which might help. I guess it's quite tedious to do, because someone has to annotate training data, which is a pretty unpleasant task.
\nAdditionally, we know that some of the images don't actually show images, but roots. At least those images are relatively easy to separate out from the ones that primarily show leaves, so it should be possible to treat those differently.
\nIn principle though, this seems like a segmentation taks that something like U-Net should do a prety good job at.
", "votes": null, "replies": [ { "id": 1150954, "author_name": "wantsu", "author_url": "", "post_date": "01/13/2021 02:07:11", "content": "Hello, Björn. Thanks for your opinion. Segmenting images is tedious in this setting , like you said. what's more, backgound information is useful for image classification (NOISE OR SIGNAL: THE ROLE OF IMAGE BACKGROUNDS IN OBJECT RECOGNITION). I was thinking that to help model focus on saliency object, less consider the backgounds.
", "votes": null, "replies": [] }, { "id": 1151132, "author_name": "zhangeng", "author_url": "", "post_date": "01/13/2021 06:44:24", "content": "I think the background is quite noisy. We can use instance segmentation to segment the important areas, but we need unsupervised training. Of course, we don't need much precision. We roughly segment the important feature areas, and then add the attention element to identify the specific categories more accurately. But we also need to do this synchronously when predicting. So there are at least three problems that are not easy to solve
\nThe program is too time-consuming, and it may not be more accurate by simple classification + fusion?
Label problem cannot be solved, so unsupervised learning is needed
Unsupervised instance segmentation. I don't know much about this field. I don't know if there is a suitable network?
Another point is that even if you are more accurate than the final test set, there are some inaccurate labels on the test set, which also leads to some differences in the final score
", "votes": null, "replies": [] } ] }, { "id": 1151622, "author_name": "saikumargv", "author_url": "", "post_date": "01/13/2021 13:08:11", "content": "A simple center crop with a reasonable cropping size might also work.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1149723": "Hello guys! \nI noticed that many training image has a **clutterred background**(The training image is collected from a wild environment), and the surrounding plants in some images are not Cassava Leaf. Based on such observation, i reckon that **the contextual information may lead to a biased model**, especially when we use random crop to augment trianing image. In this setting, croped image may lost important discriminative feature of Cassava leaf but fill with many noised contextual information.", "1149735": "Therefore, **attention** mode is needed. A model with **attention** function will help to better identify the characteristics of cassava and then classify", "1149752": "totally, agree", "1150154": "One could segment out just the cassava leaves, which might help. I guess it's quite tedious to do, because someone has to annotate training data, which is a pretty unpleasant task.\n\nAdditionally, [we know](https://www.kaggle.com/bjoernholzhauer/cassava-leaf-disease-classif-eda-cv-strategy) that some of the images don't actually show images, but roots. At least those images are relatively easy to separate out from the ones that primarily show leaves, so it should be possible to treat those differently.\n\nIn principle though, this seems like a segmentation taks that something like U-Net should do a prety good job at.", "1150954": "Hello, Björn. Thanks for your opinion. Segmenting images is tedious in this setting , like you said. what's more, backgound information is useful for image classification ([NOISE OR SIGNAL: THE ROLE OF IMAGE BACKGROUNDS IN OBJECT RECOGNITION](https://openreview.net/pdf?id=gl3D-xY7wLq)). I was thinking that to help model focus on saliency object, less consider the backgounds.", "1151132": "I think the background is quite noisy. We can use instance segmentation to segment the important areas, but we need unsupervised training. Of course, we don't need much precision. We roughly segment the important feature areas, and then add the attention element to identify the specific categories more accurately. But we also need to do this synchronously when predicting. So there are at least three problems that are not easy to solve\n\n1. The program is too time-consuming, and it may not be more accurate by simple classification + fusion?\n\n2. Label problem cannot be solved, so unsupervised learning is needed\n\n3. Unsupervised instance segmentation. I don't know much about this field. I don't know if there is a suitable network?\n\nAnother point is that even if you are more accurate than the final test set, there are some inaccurate labels on the test set, which also leads to some differences in the final score", "1151622": "A simple center crop with a reasonable cropping size might also work." }, "source": "meta" }