Hi Kagglers !
\nI have to admit I'm a bit skeptical about which strategy to adopt on that competition. It's a very new format and not easy to get started with.
\nWhat would do you say of object detection techniques ? It would mean using the images directly and converting to a full computer vision problem. Here are some of the ressources I have explored on the subject :
\nDo you have any other strategies ? I would be very interested in hearing any suggestion !
", "messages": [ { "id": "1084155", "postDate": "11/19/2020 19:40:33", "content": "Hi Kagglers !
\nI have to admit I'm a bit skeptical about which strategy to adopt on that competition. It's a very new format and not easy to get started with.
\nWhat would do you say of object detection techniques ? It would mean using the images directly and converting to a full computer vision problem. Here are some of the ressources I have explored on the subject :
\nDo you have any other strategies ? I would be very interested in hearing any suggestion !
", "rawMarkdown": "Hi Kagglers !\n\nI have to admit I'm a bit skeptical about which strategy to adopt on that competition. It's a very new format and not easy to get started with. \n\nWhat would do you say of object detection techniques ? It would mean using the images directly and converting to a full computer vision problem. Here are some of the ressources I have explored on the subject :\n- [Attention-Based Deep Neural Networks](https://arxiv.org/ftp/arxiv/papers/1811/1811.08513.pdf) ;\n- [Detecting Cancer Metastases on Gigapixel Pathology Images](https://arxiv.org/pdf/1703.02442v2.pdf) : this paper seems to be one of the most popular on the subject since it has several github repos available [here](https://github.com/Reemr/Cancer-detection), [here](https://github.com/olahosa/adl_cancer_detection), [here ](https://github.com/Srinidhi-kv/Cancer_metastasis_detection), even [here ](https://github.com/kira-95/adl_cancer_detection) and finally [here](https://github.com/virabehnam/CAMELYON16-Breast-Cancer-Detection).\n- [DeepSlide: A Sliding Window Framework for Classification of High Resolution Microscopy Images](https://github.com/BMIRDS/deepslide).\n\nDo you have any other strategies ? I would be very interested in hearing any suggestion !", "votes": null }, { "id": "1084212", "postDate": "11/19/2020 20:42:23", "content": "This my first competition where the task is segmentation, i search a little and i found Graph U-Nets and an implementation with pytorch geometric maybe could be interesting for you. I read about graph neural networks and saw that they are used in papers where domain knowledge gives and advantage.
", "rawMarkdown": "This my first competition where the task is **segmentation**, i search a little and i found [Graph U-Nets](https://arxiv.org/pdf/1905.05178.pdf) and an implementation with [pytorch geometric](https://pytorchgeometric.readthedocs.io/en/latest/modules/nn.html#torch_geometric.nn.models.GraphUNet) maybe could be interesting for you. I read about graph neural networks and saw that they are used in papers where **domain knowledge** gives and advantage.", "votes": null }, { "id": "1084831", "postDate": "11/20/2020 12:36:48", "content": "Thanks for the suggestions π
", "rawMarkdown": "Thanks for the suggestions π", "votes": null }, { "id": "1085206", "postDate": "11/20/2020 18:48:25", "content": "To offer an alternative to standard segmentation approaches: you can try classifying small image patches, and then thresholding your inference results (or some other secondary processing step) to create a region mask.
\nAdvantages and disadvantages to each approach. I'm planning on trying both to see which works better!
", "rawMarkdown": "To offer an alternative to standard segmentation approaches: you can try classifying small image patches, and then thresholding your inference results (or some other secondary processing step) to create a region mask. \n\nAdvantages and disadvantages to each approach. I'm planning on trying both to see which works better!", "votes": null }, { "id": "1094849", "postDate": "11/29/2020 02:20:24", "content": "This my first competition where the task is segmentation, tooβ¦. I can figure out how many classes are here, One? Two? when doing training. (H x W x C) where C is the total number of classes.
\nThanks.
", "rawMarkdown": "This my first competition where the task is segmentation, too.... I can figure out how many classes are here, One? Two? when doing training. (H x W x C) where C is the total number of classes.\n\nThanks.", "votes": null } ], "comments": [ { "id": 1084212, "author_name": "hiramcho", "author_url": "", "post_date": "11/19/2020 20:42:23", "content": "This my first competition where the task is segmentation, i search a little and i found Graph U-Nets and an implementation with pytorch geometric maybe could be interesting for you. I read about graph neural networks and saw that they are used in papers where domain knowledge gives and advantage.
", "votes": null, "replies": [ { "id": 1084831, "author_name": "louise2001", "author_url": "", "post_date": "11/20/2020 12:36:48", "content": "Thanks for the suggestions π
", "votes": null, "replies": [] } ] }, { "id": 1085206, "author_name": "paulgamble", "author_url": "", "post_date": "11/20/2020 18:48:25", "content": "To offer an alternative to standard segmentation approaches: you can try classifying small image patches, and then thresholding your inference results (or some other secondary processing step) to create a region mask.
\nAdvantages and disadvantages to each approach. I'm planning on trying both to see which works better!
", "votes": null, "replies": [] }, { "id": 1094849, "author_name": "oscarrangel", "author_url": "", "post_date": "11/29/2020 02:20:24", "content": "This my first competition where the task is segmentation, tooβ¦. I can figure out how many classes are here, One? Two? when doing training. (H x W x C) where C is the total number of classes.
\nThanks.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1084155": "Hi Kagglers !\n\nI have to admit I'm a bit skeptical about which strategy to adopt on that competition. It's a very new format and not easy to get started with. \n\nWhat would do you say of object detection techniques ? It would mean using the images directly and converting to a full computer vision problem. Here are some of the ressources I have explored on the subject :\n- [Attention-Based Deep Neural Networks](https://arxiv.org/ftp/arxiv/papers/1811/1811.08513.pdf) ;\n- [Detecting Cancer Metastases on Gigapixel Pathology Images](https://arxiv.org/pdf/1703.02442v2.pdf) : this paper seems to be one of the most popular on the subject since it has several github repos available [here](https://github.com/Reemr/Cancer-detection), [here](https://github.com/olahosa/adl_cancer_detection), [here ](https://github.com/Srinidhi-kv/Cancer_metastasis_detection), even [here ](https://github.com/kira-95/adl_cancer_detection) and finally [here](https://github.com/virabehnam/CAMELYON16-Breast-Cancer-Detection).\n- [DeepSlide: A Sliding Window Framework for Classification of High Resolution Microscopy Images](https://github.com/BMIRDS/deepslide).\n\nDo you have any other strategies ? I would be very interested in hearing any suggestion !", "1084212": "This my first competition where the task is **segmentation**, i search a little and i found [Graph U-Nets](https://arxiv.org/pdf/1905.05178.pdf) and an implementation with [pytorch geometric](https://pytorchgeometric.readthedocs.io/en/latest/modules/nn.html#torch_geometric.nn.models.GraphUNet) maybe could be interesting for you. I read about graph neural networks and saw that they are used in papers where **domain knowledge** gives and advantage.", "1084831": "Thanks for the suggestions π", "1085206": "To offer an alternative to standard segmentation approaches: you can try classifying small image patches, and then thresholding your inference results (or some other secondary processing step) to create a region mask. \n\nAdvantages and disadvantages to each approach. I'm planning on trying both to see which works better!", "1094849": "This my first competition where the task is segmentation, too.... I can figure out how many classes are here, One? Two? when doing training. (H x W x C) where C is the total number of classes.\n\nThanks." }, "source": "meta" }