One of the key issues in this competition is weak supervision. I have known some solutions of other people, one of which is to assign multiple labels to each cell in the image when the image corresponds to multiple labels, and perform instance segmentation model training, and I think this approach will Introduce some noise. The other is to directly discard images with multiple labels, directly use single-label images, and generate labels after segmentation.
\nMy thought is whether we can segment the cells first. For images with only single labels, we can directly assign labels, while images with multiple labels will not be used in the first training stage. Use a single-label image to train the feature extractor and classifier, and then perform feature extraction on each segmented cell in the multi-label image, and take the same number of cluster centers as the number of labels for clustering, for example, use kmeans. Then, the single-label image classifier is used to classify and pseudo-label this category. In this way, no matter single-label image or multi-label image, cell-level classification labels can be obtained.
\nBecause I am new to kaggle, I don't know whether this idea is reasonable, I hope everyone can provide me with suggestions, thank you.
", "messages": [ { "id": 1270939, "postDate": "2021-04-12T07:09:58.890Z", "content": "One of the key issues in this competition is weak supervision. I have known some solutions of other people, one of which is to assign multiple labels to each cell in the image when the image corresponds to multiple labels, and perform instance segmentation model training, and I think this approach will Introduce some noise. The other is to directly discard images with multiple labels, directly use single-label images, and generate labels after segmentation.
\nMy thought is whether we can segment the cells first. For images with only single labels, we can directly assign labels, while images with multiple labels will not be used in the first training stage. Use a single-label image to train the feature extractor and classifier, and then perform feature extraction on each segmented cell in the multi-label image, and take the same number of cluster centers as the number of labels for clustering, for example, use kmeans. Then, the single-label image classifier is used to classify and pseudo-label this category. In this way, no matter single-label image or multi-label image, cell-level classification labels can be obtained.
\nBecause I am new to kaggle, I don't know whether this idea is reasonable, I hope everyone can provide me with suggestions, thank you.
", "rawMarkdown": "One of the key issues in this competition is weak supervision. I have known some solutions of other people, one of which is to assign multiple labels to each cell in the image when the image corresponds to multiple labels, and perform instance segmentation model training, and I think this approach will Introduce some noise. The other is to directly discard images with multiple labels, directly use single-label images, and generate labels after segmentation.\n\nMy thought is whether we can segment the cells first. For images with only single labels, we can directly assign labels, while images with multiple labels will not be used in the first training stage. Use a single-label image to train the feature extractor and classifier, and then perform feature extraction on each segmented cell in the multi-label image, and take the same number of cluster centers as the number of labels for clustering, for example, use kmeans. Then, the single-label image classifier is used to classify and pseudo-label this category. In this way, no matter single-label image or multi-label image, cell-level classification labels can be obtained.\n\nBecause I am new to kaggle, I don't know whether this idea is reasonable, I hope everyone can provide me with suggestions, thank you.", "votes": 4 }, { "id": 1271107, "postDate": "2021-04-12T09:57:51.667Z", "content": "Trial and error. And pay attention to time and memory.
", "rawMarkdown": "Trial and error. And pay attention to time and memory." } ], "comments": [ { "id": 1271107, "author_name": "Alien", "author_url": "", "post_date": "2021-04-12T09:57:51.667000", "content": "Trial and error. And pay attention to time and memory.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1270939": "One of the key issues in this competition is weak supervision. I have known some solutions of other people, one of which is to assign multiple labels to each cell in the image when the image corresponds to multiple labels, and perform instance segmentation model training, and I think this approach will Introduce some noise. The other is to directly discard images with multiple labels, directly use single-label images, and generate labels after segmentation.\n\nMy thought is whether we can segment the cells first. For images with only single labels, we can directly assign labels, while images with multiple labels will not be used in the first training stage. Use a single-label image to train the feature extractor and classifier, and then perform feature extraction on each segmented cell in the multi-label image, and take the same number of cluster centers as the number of labels for clustering, for example, use kmeans. Then, the single-label image classifier is used to classify and pseudo-label this category. In this way, no matter single-label image or multi-label image, cell-level classification labels can be obtained.\n\nBecause I am new to kaggle, I don't know whether this idea is reasonable, I hope everyone can provide me with suggestions, thank you.", "1271107": "Trial and error. And pay attention to time and memory." } }