![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F21ea0caa4e1fed2f9f8a11fa4deeed8a%2FSelection_074.png?generation=1607184759886359&alt=media\")
I made a few quizzes for you to understand the data more.
\nyou should use these to decide how to augment the data, post/pre-processing, remove bias from train/test , label cleaning, estimate byes error (if you assume byes error = human error) etc …
\nusing perception similarity, you may want to think about metric learning, etc
\nplease see the questions below:
", "messages": [ { "id": "1103085", "postDate": "12/05/2020 16:11:03", "content": "I made a few quizzes for you to understand the data more.
\nyou should use these to decide how to augment the data, post/pre-processing, remove bias from train/test , label cleaning, estimate byes error (if you assume byes error = human error) etc …
\nusing perception similarity, you may want to think about metric learning, etc
\nplease see the questions below:
", "rawMarkdown": "I made a few quizzes for you to understand the data more.\n\nyou should use these to decide how to augment the data, post/pre-processing, remove bias from train/test , label cleaning, estimate byes error (if you assume byes error = human error) etc ...\n\nusing perception similarity, you may want to think about metric learning, etc\n\n---\n \nplease see the questions below:", "votes": null }, { "id": "1103089", "postDate": "12/05/2020 16:13:05", "content": "question.1
\nthe white dots are centers of given kaggle annotations in train dataset.
\nthere are three missing ones. can you spot them?
\n(answer are in the attachment)
question.2
\nThere is one missing annotation. where is it?
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fdc67a2f1425f2c4e7c6aff6744880d59%2FSelection_076.png?generation=1607185321196186&alt=media\")
A1: Yep at top
", "rawMarkdown": "A1: Yep at top", "votes": null }, { "id": "1103099", "postDate": "12/05/2020 16:26:50", "content": "A2: I think at the bottom corner edges
", "rawMarkdown": "A2: I think at the bottom corner edges", "votes": null }, { "id": "1103106", "postDate": "12/05/2020 16:34:03", "content": "answer at the attached file (click Selection_075.png and Selection_077.png)
", "rawMarkdown": "answer at the attached file (click Selection_075.png and Selection_077.png)", "votes": null }, { "id": "1103126", "postDate": "12/05/2020 16:41:41", "content": "question.3
\none of the white dot annotations is wrong ( not consistent with kaggle ground truth)
\nwhich one is it?
\n(answer are in the attachment)
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fbdc5d3c0900cee83917ed3b167c1f619%2FSelection_078.png?generation=1607186478900923&alt=media\")
how many white dot annotations are wrong?
\n(answer are in the attachment)
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc1d6d066a96328643313d2ca4ed85656%2FSelection_080.png?generation=1607187118619593&alt=media\")
there are one wrong annotation and two missing ones. can you spot them?
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4492b393a3870f28b40d22adb8988275%2FSelection_082.png?generation=1607187861165326&alt=media\")
Okay i will make a try.
\nBut have you solved the issue of allocating more memory in inference step
I estimate outlier (or human error annotation?) in train is about 5%. if you assume average dice score per instance is about 0.95, then the best achievable score is about 0.95*0.95 = 0.9025
\nas for the public test images, they are very close in terms of color and texture to the train images.
\nhence validation results should be closed the LB score.
now i am only worried about the annotation in the following regions:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fce5ec334db0ef086bb8281d92628de88%2FSelection_085.png?generation=1607189253421414&alt=media\")
the concern is NOT if there are glomerulus or not, but rather if they are annotated or not.
\n(one may not annotate them because they are not useful)
one can estimate the number of glomerulus using the density glomerulus of nearby region.
\nTo know if they are annotated, I can only think of probing the server. (e.g. check drop of LB score by removing any prediction in the detected bright or dark region)
not yet. My flow of work is:
\nhi @hengck23, how should I verify my answers are correct? I clicked the images above but nothing happened.
", "rawMarkdown": "hi @hengck23, how should I verify my answers are correct? I clicked the images above but nothing happened.", "votes": null }, { "id": "1117072", "postDate": "12/17/2020 17:48:20", "content": "click on link of the attached file, e.g. https://storage.googleapis.com/kaggle-forum-message-attachments/1103126/17617/Selection_079.png
", "rawMarkdown": "click on link of the attached file, e.g. https://storage.googleapis.com/kaggle-forum-message-attachments/1103126/17617/Selection_079.png", "votes": null } ], "comments": [ { "id": 1103089, "author_name": "hengck23", "author_url": "", "post_date": "12/05/2020 16:13:05", "content": "question.1
\nthe white dots are centers of given kaggle annotations in train dataset.
\nthere are three missing ones. can you spot them?
\n(answer are in the attachment)
question.2
\nThere is one missing annotation. where is it?
A1: Yep at top
", "votes": null, "replies": [] }, { "id": 1103099, "author_name": "morizin", "author_url": "", "post_date": "12/05/2020 16:26:50", "content": "A2: I think at the bottom corner edges
", "votes": null, "replies": [] }, { "id": 1103106, "author_name": "hengck23", "author_url": "", "post_date": "12/05/2020 16:34:03", "content": "answer at the attached file (click Selection_075.png and Selection_077.png)
", "votes": null, "replies": [] }, { "id": 1103159, "author_name": "morizin", "author_url": "", "post_date": "12/05/2020 17:14:25", "content": "Okay i will make a try.
\nBut have you solved the issue of allocating more memory in inference step
not yet. My flow of work is:
\nhi @hengck23, how should I verify my answers are correct? I clicked the images above but nothing happened.
", "votes": null, "replies": [] }, { "id": 1117072, "author_name": "hengck23", "author_url": "", "post_date": "12/17/2020 17:48:20", "content": "click on link of the attached file, e.g. https://storage.googleapis.com/kaggle-forum-message-attachments/1103126/17617/Selection_079.png
", "votes": null, "replies": [] } ] }, { "id": 1103126, "author_name": "hengck23", "author_url": "", "post_date": "12/05/2020 16:41:41", "content": "question.3
\none of the white dot annotations is wrong ( not consistent with kaggle ground truth)
\nwhich one is it?
\n(answer are in the attachment)
how many white dot annotations are wrong?
\n(answer are in the attachment)
there are one wrong annotation and two missing ones. can you spot them?
\n
I estimate outlier (or human error annotation?) in train is about 5%. if you assume average dice score per instance is about 0.95, then the best achievable score is about 0.95*0.95 = 0.9025
\nas for the public test images, they are very close in terms of color and texture to the train images.
\nhence validation results should be closed the LB score.
now i am only worried about the annotation in the following regions:
\n
the concern is NOT if there are glomerulus or not, but rather if they are annotated or not.
\n(one may not annotate them because they are not useful)
one can estimate the number of glomerulus using the density glomerulus of nearby region.
\nTo know if they are annotated, I can only think of probing the server. (e.g. check drop of LB score by removing any prediction in the detected bright or dark region)