I expected some shift because I knew the resolution was different for the private LB but I definitely did not expect this. This makes me that much more curious for Clevert and Lumos517's solution which managed to stay strong for both leaderboards. Also an impressive performance from @ojimaryoji who moved up 1053 places!!!! It makes me think that some unique strategy preformed very similar in both public and private LB in terms of score and then it just paid off with the shift? Congrats to all, I learned so much!
", "messages": [ { "id": 2640441, "postDate": "2024-02-07T00:11:47.123Z", "content": "I expected some shift because I knew the resolution was different for the private LB but I definitely did not expect this. This makes me that much more curious for Clevert and Lumos517's solution which managed to stay strong for both leaderboards. Also an impressive performance from @ojimaryoji who moved up 1053 places!!!! It makes me think that some unique strategy preformed very similar in both public and private LB in terms of score and then it just paid off with the shift? Congrats to all, I learned so much!
", "rawMarkdown": "I expected some shift because I knew the resolution was different for the private LB but I definitely did not expect this. This makes me that much more curious for Clevert and Lumos517's solution which managed to stay strong for both leaderboards. Also an impressive performance from @ojimaryoji who moved up 1053 places!!!! It makes me think that some unique strategy preformed very similar in both public and private LB in terms of score and then it just paid off with the shift? Congrats to all, I learned so much!", "votes": 4 }, { "id": 2640444, "postDate": "2024-02-07T00:14:16.677Z", "content": "This was my first for computer vision project, and I joined the competition quite late (two weeks before the deadline!). While my highest ranking was 122, it unfortunately dropped to 318 due to inconsistencies between private and public scores.
\nDespite the rank drop, I gained valuable insights thanks to the shared kernels (Special thanks to @misakimatsutomo). Congratulations to everyone who achieved great results!
\nI've shared my notebook for anyone interested: [Infer] Segmentation Mask.
\nI'd be happy to receive any feedback!
", "rawMarkdown": "This was my first for computer vision project, and I joined the competition quite late (two weeks before the deadline!). While my highest ranking was 122, it unfortunately dropped to 318 due to inconsistencies between private and public scores. \n\nDespite the rank drop, I gained valuable insights thanks to the shared kernels (Special thanks to @misakimatsutomo). Congratulations to everyone who achieved great results!\n\nI've shared my notebook for anyone interested: [[Infer] Segmentation Mask](https://www.kaggle.com/code/minhsienweng/infer-segmentation-mask).\n\nI'd be happy to receive any feedback!", "votes": 1, "replies": [ { "id": 2640446, "postDate": "2024-02-07T00:15:46.527Z", "content": "Yes absolutely, and I cannot wait to see what they settled on for submission!
", "rawMarkdown": "Yes absolutely, and I cannot wait to see what they settled on for submission!", "votes": 1 } ] }, { "id": 2640462, "postDate": "2024-02-07T00:29:45.530Z", "content": "maybe threshold is matter, larger threshold lead to higher private score
", "rawMarkdown": "maybe threshold is matter, larger threshold lead to higher private score", "votes": 2, "replies": [ { "id": 2640471, "postDate": "2024-02-07T00:32:26.323Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 2640460, "postDate": "2024-02-07T00:29:27.877Z", "content": "\"This makes me that much more curious for Clevert and Lumos517's solution which managed to stay strong\"
\ni would guess one factor is better scale handling via data augmentation and modelling. let's wait and see :)
", "rawMarkdown": "\"This makes me that much more curious for Clevert and Lumos517's solution which managed to stay strong\"\n\n\ni would guess one factor is better scale handling via data augmentation and modelling. let's wait and see :)", "votes": 2 }, { "id": 2640453, "postDate": "2024-02-07T00:24:15.523Z", "content": "I used the popular code. Found a lot of problems. I spent last two days mostly rewrite the code. I run out of time. Can't quite rewrite everything. The code would failed big time if there are more than two kidneys. Also it used kidney 5 image size for all the kidneys. It just happen kedney 5 and 6 have the same xy size. I found out th_percentile was the problem. I discovered it relatively late. It was much higher for my local run for kidney 3. My local kidney 3 (only used the part of sections) was around th_percentile 0.04. My best score used th_percentile=0.0173 instead of 0.0143, even though it scored lower at public LB. Should have used higher. But there was no way for me to know what could have been the best as public LB was not reliable. Tried to rewrite it to use absolute threshold. I encounter some bugs that I couldn't solve it before the deadline.
", "rawMarkdown": "I used the popular code. Found a lot of problems. I spent last two days mostly rewrite the code. I run out of time. Can't quite rewrite everything. The code would failed big time if there are more than two kidneys. Also it used kidney 5 image size for all the kidneys. It just happen kedney 5 and 6 have the same xy size. I found out th_percentile was the problem. I discovered it relatively late. It was much higher for my local run for kidney 3. My local kidney 3 (only used the part of sections) was around th_percentile 0.04. My best score used th_percentile=0.0173 instead of 0.0143, even though it scored lower at public LB. Should have used higher. But there was no way for me to know what could have been the best as public LB was not reliable. Tried to rewrite it to use absolute threshold. I encounter some bugs that I couldn't solve it before the deadline. ", "votes": 2, "replies": [ { "id": 2640455, "postDate": "2024-02-07T00:26:45.943Z", "content": "Interesting! So this was a form of the public notebook you started with? If so very good job!
", "rawMarkdown": "Interesting! So this was a form of the public notebook you started with? If so very good job!", "votes": 1, "replies": [ { "id": 2640482, "postDate": "2024-02-07T00:37:52.863Z", "content": "I used different code, partly from others posted here with my own code. It scored 0.65 top on public LB. I used the popular code about a week ago. I think the tile approach was good. But th_percentile was guess work.
", "rawMarkdown": "I used different code, partly from others posted here with my own code. It scored 0.65 top on public LB. I used the popular code about a week ago. I think the tile approach was good. But th_percentile was guess work. ", "votes": 1 } ] } ] } ], "comments": [ { "id": 2640444, "author_name": "Min-Hsien Weng", "author_url": "", "post_date": "2024-02-07T00:14:16.677000", "content": "This was my first for computer vision project, and I joined the competition quite late (two weeks before the deadline!). While my highest ranking was 122, it unfortunately dropped to 318 due to inconsistencies between private and public scores.
\nDespite the rank drop, I gained valuable insights thanks to the shared kernels (Special thanks to @misakimatsutomo). Congratulations to everyone who achieved great results!
\nI've shared my notebook for anyone interested: [Infer] Segmentation Mask.
\nI'd be happy to receive any feedback!
", "votes": 1, "replies": [ { "id": 2640446, "author_name": "Cody_Null", "author_url": "", "post_date": "2024-02-07T00:15:46.527000", "content": "Yes absolutely, and I cannot wait to see what they settled on for submission!
", "votes": 1, "replies": [] } ] }, { "id": 2640462, "author_name": "tanxxx", "author_url": "", "post_date": "2024-02-07T00:29:45.530000", "content": "maybe threshold is matter, larger threshold lead to higher private score
", "votes": 2, "replies": [ { "id": 2640471, "author_name": "", "author_url": "", "post_date": "2024-02-07T00:32:26.323000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 2640460, "author_name": "hengck23", "author_url": "", "post_date": "2024-02-07T00:29:27.877000", "content": "\"This makes me that much more curious for Clevert and Lumos517's solution which managed to stay strong\"
\ni would guess one factor is better scale handling via data augmentation and modelling. let's wait and see :)
", "votes": 2, "replies": [] }, { "id": 2640453, "author_name": "joejeo1", "author_url": "", "post_date": "2024-02-07T00:24:15.523000", "content": "I used the popular code. Found a lot of problems. I spent last two days mostly rewrite the code. I run out of time. Can't quite rewrite everything. The code would failed big time if there are more than two kidneys. Also it used kidney 5 image size for all the kidneys. It just happen kedney 5 and 6 have the same xy size. I found out th_percentile was the problem. I discovered it relatively late. It was much higher for my local run for kidney 3. My local kidney 3 (only used the part of sections) was around th_percentile 0.04. My best score used th_percentile=0.0173 instead of 0.0143, even though it scored lower at public LB. Should have used higher. But there was no way for me to know what could have been the best as public LB was not reliable. Tried to rewrite it to use absolute threshold. I encounter some bugs that I couldn't solve it before the deadline.
", "votes": 2, "replies": [ { "id": 2640455, "author_name": "Cody_Null", "author_url": "", "post_date": "2024-02-07T00:26:45.943000", "content": "Interesting! So this was a form of the public notebook you started with? If so very good job!
", "votes": 1, "replies": [ { "id": 2640482, "author_name": "joejeo1", "author_url": "", "post_date": "2024-02-07T00:37:52.863000", "content": "I used different code, partly from others posted here with my own code. It scored 0.65 top on public LB. I used the popular code about a week ago. I think the tile approach was good. But th_percentile was guess work.
", "votes": 1, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2640441": "I expected some shift because I knew the resolution was different for the private LB but I definitely did not expect this. This makes me that much more curious for Clevert and Lumos517's solution which managed to stay strong for both leaderboards. Also an impressive performance from @ojimaryoji who moved up 1053 places!!!! It makes me think that some unique strategy preformed very similar in both public and private LB in terms of score and then it just paid off with the shift? Congrats to all, I learned so much!", "2640444": "This was my first for computer vision project, and I joined the competition quite late (two weeks before the deadline!). While my highest ranking was 122, it unfortunately dropped to 318 due to inconsistencies between private and public scores. \n\nDespite the rank drop, I gained valuable insights thanks to the shared kernels (Special thanks to @misakimatsutomo). Congratulations to everyone who achieved great results!\n\nI've shared my notebook for anyone interested: [[Infer] Segmentation Mask](https://www.kaggle.com/code/minhsienweng/infer-segmentation-mask).\n\nI'd be happy to receive any feedback!", "2640462": "maybe threshold is matter, larger threshold lead to higher private score", "2640460": "\"This makes me that much more curious for Clevert and Lumos517's solution which managed to stay strong\"\n\n\ni would guess one factor is better scale handling via data augmentation and modelling. let's wait and see :)", "2640453": "I used the popular code. Found a lot of problems. I spent last two days mostly rewrite the code. I run out of time. Can't quite rewrite everything. The code would failed big time if there are more than two kidneys. Also it used kidney 5 image size for all the kidneys. It just happen kedney 5 and 6 have the same xy size. I found out th_percentile was the problem. I discovered it relatively late. It was much higher for my local run for kidney 3. My local kidney 3 (only used the part of sections) was around th_percentile 0.04. My best score used th_percentile=0.0173 instead of 0.0143, even though it scored lower at public LB. Should have used higher. But there was no way for me to know what could have been the best as public LB was not reliable. Tried to rewrite it to use absolute threshold. I encounter some bugs that I couldn't solve it before the deadline. " } }