![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F479538%2F982dcbd71feef057a74d75a08696cbf2%2F095bf7a1f_bf_correction.jpg?generation=1607320094167381&alt=media\")
@addisonhoward @leahscherschel
\nAs @hengck23 have pointed out, some of the masks seem to be off.
\nhttps://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/201816
The image below represents the mask in figure 095bf7a1f, which he also pointed out.
The mask in the above figure seems off to me, and in fact the CV of this figure is not very good. If we shift this figure by 25 horizontally and 15 vertically, it looks like the figure below.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F479538%2Fa53c2b19ea0ca65a54c5dcf2e145e361%2F095bf7a1f_af_correction.jpg?generation=1607320377057733&alt=media\")
This seems better, but are we allowed to fix the mask ourselves? If this is allowed, then the next question is whether we can get unlabeled external data and apply minor modifications ourselves according to the predicted masks.
\nWhat is allowed and what is not?
\nI personally suppose that if this is not cleared, the competition can become opaque and unfair.
", "messages": [ { "id": "1104620", "postDate": "12/07/2020 06:05:02", "content": "@addisonhoward @leahscherschel
\nAs @hengck23 have pointed out, some of the masks seem to be off.
\nhttps://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/201816
The image below represents the mask in figure 095bf7a1f, which he also pointed out.
The mask in the above figure seems off to me, and in fact the CV of this figure is not very good. If we shift this figure by 25 horizontally and 15 vertically, it looks like the figure below.
\n
This seems better, but are we allowed to fix the mask ourselves? If this is allowed, then the next question is whether we can get unlabeled external data and apply minor modifications ourselves according to the predicted masks.
\nWhat is allowed and what is not?
\nI personally suppose that if this is not cleared, the competition can become opaque and unfair.
", "rawMarkdown": "addisonhoward @leahscherschel \n\nAs @hengck23 have pointed out, some of the masks seem to be off. \nhttps://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/201816\n\nThe image below represents the mask in figure `095bf7a1f`, which he also pointed out.\n\n\n\nThe mask in the above figure seems off to me, and in fact the CV of this figure is not very good. If we shift this figure by **25 horizontally** and **15 vertically**, it looks like the figure below.\n\n\n\nThis seems better, but are we allowed to fix the mask ourselves? If this is allowed, then the next question is whether we can get unlabeled external data and apply minor modifications ourselves according to the predicted masks.\n\nWhat is allowed and what is not?\n\nI personally suppose that if this is not cleared, the competition can become opaque and unfair.", "votes": null }, { "id": "1104627", "postDate": "12/07/2020 06:12:15", "content": "actually, the main concern is the test set. we hope the test ground truth at the server is not shifted.
\nfor train set, you are free to do what you want.
\n\"25 horizontally and 15 vertically,\"
\nthe best way to determine this is to write a program to do brute force search that will max your CV for a variety of model and flip images. I think manually shifting (or judgment by eye) is not optimum.
Thank you for your comment.
\n\n\nfor train set, you are free to do what you want.
\n
Does \"the train data\" you refer to include external data? Is it okay to hand label for external data?
", "rawMarkdown": "hengck23 \n\nThank you for your comment.\n\n>for train set, you are free to do what you want.\n\nDoes \"the train data\" you refer to include external data? Is it okay to hand label for external data?", "votes": null }, { "id": "1104701", "postDate": "12/07/2020 07:20:54", "content": "you can label external data
", "rawMarkdown": "you can label external data", "votes": null }, { "id": "1112125", "postDate": "12/14/2020 10:00:20", "content": "@hengck23 thanks for your comment.
\n I dont understand why flip can help us find out the shifted images.
\nIn my opinion,whenwe flip images,mask need flip together.
\nWould your explain more about it ?
\nThx.
\n", "rawMarkdown": "hengck23 thanks for your comment.\n I dont understand why flip can help us find out the shifted images.\nIn my opinion,whenwe flip images,mask need flip together.\nWould your explain more about it ?\nThx.\n> the best way to determine this is to write a program to do brute force search that will max your CV for a variety of model and flip images.", "votes": null } ], "comments": [ { "id": 1104627, "author_name": "hengck23", "author_url": "", "post_date": "12/07/2020 06:12:15", "content": "the best way to determine this is to write a program to do brute force search that will max your CV for a variety of model and flip images.
\n
actually, the main concern is the test set. we hope the test ground truth at the server is not shifted.
\nfor train set, you are free to do what you want.
\n\"25 horizontally and 15 vertically,\"
\nthe best way to determine this is to write a program to do brute force search that will max your CV for a variety of model and flip images. I think manually shifting (or judgment by eye) is not optimum.
Thank you for your comment.
\n\n\nfor train set, you are free to do what you want.
\n
Does \"the train data\" you refer to include external data? Is it okay to hand label for external data?
", "votes": null, "replies": [] }, { "id": 1104701, "author_name": "hengck23", "author_url": "", "post_date": "12/07/2020 07:20:54", "content": "you can label external data
", "votes": null, "replies": [] } ] }, { "id": 1112125, "author_name": "mashepburn8", "author_url": "", "post_date": "12/14/2020 10:00:20", "content": "@hengck23 thanks for your comment.
\n I dont understand why flip can help us find out the shifted images.
\nIn my opinion,whenwe flip images,mask need flip together.
\nWould your explain more about it ?
\nThx.
\n", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1104620": "addisonhoward @leahscherschel \n\nAs @hengck23 have pointed out, some of the masks seem to be off. \nhttps://www.kaggle.com/c/hubmap-kidney-segmentation/discussion/201816\n\nThe image below represents the mask in figure `095bf7a1f`, which he also pointed out.\n\n\n\nThe mask in the above figure seems off to me, and in fact the CV of this figure is not very good. If we shift this figure by **25 horizontally** and **15 vertically**, it looks like the figure below.\n\n\n\nThis seems better, but are we allowed to fix the mask ourselves? If this is allowed, then the next question is whether we can get unlabeled external data and apply minor modifications ourselves according to the predicted masks.\n\nWhat is allowed and what is not?\n\nI personally suppose that if this is not cleared, the competition can become opaque and unfair.", "1104627": "actually, the main concern is the test set. we hope the test ground truth at the server is not shifted.\n\nfor train set, you are free to do what you want.\n\n\"25 horizontally and 15 vertically,\"\nthe best way to determine this is to write a program to do brute force search that will max your CV for a variety of model and flip images. I think manually shifting (or judgment by eye) is not optimum.", "1104648": "hengck23 \n\nThank you for your comment.\n\n>for train set, you are free to do what you want.\n\nDoes \"the train data\" you refer to include external data? Is it okay to hand label for external data?", "1104701": "you can label external data", "1112125": "hengck23 thanks for your comment.\n I dont understand why flip can help us find out the shifted images.\nIn my opinion,whenwe flip images,mask need flip together.\nWould your explain more about it ?\nThx.\n> the best way to determine this is to write a program to do brute force search that will max your CV for a variety of model and flip images." }, "source": "meta" }the best way to determine this is to write a program to do brute force search that will max your CV for a variety of model and flip images.
\n