{ "id": 679220, "title": "18th : median filter x 7 post processing is very strong", "url": "/competitions/vesuvius-challenge-surface-detection/writeups/18th-median-filter-x-7-post-processing-is-very-s", "author_name": "", "post_date": "2026-02-28T00:43:45.663Z", "votes": 32, "comment_count": 12, "views": 0, "content": "

Great competition—thanks to the Kaggle team for organizing and running everything so smoothly!

", "messages": [ { "id": "3414934", "postDate": "02/28/2026 00:31:06", "content": "

Great competition—thanks to the Kaggle team for organizing and running everything so smoothly!

", "rawMarkdown": "Great competition—thanks to the Kaggle team for organizing and running everything so smoothly!", "votes": null }, { "id": "3414940", "postDate": "02/28/2026 01:06:27", "content": "

6 models. runtime 6h. single thread

", "rawMarkdown": "6 models. runtime 6h. single thread", "votes": null }, { "id": "3414975", "postDate": "02/28/2026 02:43:39", "content": "

code \nhttps://www.kaggle.com/code/sugupoko/v30-4-group-ensemble-blend-blur-bright-lowco?scriptVersionId=300017285

", "rawMarkdown": "code \nhttps://www.kaggle.com/code/sugupoko/v30-4-group-ensemble-blend-blur-bright-lowco?scriptVersionId=300017285", "votes": null }, { "id": "3415131", "postDate": "02/28/2026 08:15:09", "content": "

Interesting solution!

\n

Did you compress in Z every input sample, or you decided the axis of compression depending on some analysis?

", "rawMarkdown": "Interesting solution!\n\nDid you compress in Z every input sample, or you decided the axis of compression depending on some analysis?", "votes": null }, { "id": "3415146", "postDate": "02/28/2026 09:13:02", "content": "

Architecture : host model\nconfig :

\n", "rawMarkdown": "Architecture : host model\nconfig : \n- Pretrained model: 200 epochs, trained on ~700 pseudo-labeled samples + 400 external samples.\n- Global-coverage model: 500 epochs, trained without pseudo-labeled data.\n- Specialized model: 100 epochs, trained on selected data (~100–200 samples).", "votes": null }, { "id": "3415148", "postDate": "02/28/2026 09:16:15", "content": "

Thank you for your comment!!

\n

No, I didn’t choose it based on any analysis. Given the VRAM constraints of the RTX 5090, I just fixed it at around that range.

", "rawMarkdown": "Thank you for your comment!!\n\nNo, I didn’t choose it based on any analysis. Given the VRAM constraints of the RTX 5090, I just fixed it at around that range.", "votes": null }, { "id": "3415362", "postDate": "02/28/2026 19:01:10", "content": "

How? :O\nChecking your shared notebook.

", "rawMarkdown": "How? :O\nChecking your shared notebook.", "votes": null }, { "id": "3415441", "postDate": "02/28/2026 23:25:08", "content": "

Z compression is very fast. In most cases, Z is redundant.

", "rawMarkdown": "Z compression is very fast. In most cases, Z is redundant.", "votes": null }, { "id": "3415444", "postDate": "02/28/2026 23:28:41", "content": "

Congratulations on strong 18th place finish. Can you explain the \"median filter x7 times\"? I'm curious how this works. Do you ensemble models and/or use TTA and then take the median?

", "rawMarkdown": "Congratulations on strong 18th place finish. Can you explain the \"median filter x7 times\"? I'm curious how this works. Do you ensemble models and/or use TTA and then take the median?", "votes": null }, { "id": "3415448", "postDate": "02/28/2026 23:40:41", "content": "

Thank you for your comment!!!

\n

After ensemble and threshold-based binarization, I simply apply the following code.\nThe post-processing is minimal but very strong.

\n
from scipy.ndimage import median_filter\n\nPP_MEDIAN_ITER = 7\nPP_MEDIAN_SIZE = 3\n\ndef postprocess(mask: np.ndarray) -> np.ndarray:\n    mask = mask.copy()\n    for _ in range(PP_MEDIAN_ITER):\n        mask = (median_filter(mask, size=PP_MEDIAN_SIZE) > 0).astype(np.uint8)\n    return mask.astype(np.uint8)\n\nmask = probs >= THRESHOLD\nmask = postprocess(mask.astype(np.uint8))\n
\n

I think the last cell will be helpful as a reference.\nhttps://www.kaggle.com/code/sugupoko/v30-4-group-ensemble-blend-blur-bright-lowco?scriptVersionId=300348690&cellId=16

", "rawMarkdown": "Thank you for your comment!!!\n\nAfter ensemble and threshold-based binarization, I simply apply the following code.\nThe post-processing is minimal but very strong.\n\n```python\nfrom scipy.ndimage import median_filter\n\nPP_MEDIAN_ITER = 7\nPP_MEDIAN_SIZE = 3\n\ndef postprocess(mask: np.ndarray) -> np.ndarray:\n mask = mask.copy()\n for _ in range(PP_MEDIAN_ITER):\n mask = (median_filter(mask, size=PP_MEDIAN_SIZE) > 0).astype(np.uint8)\n return mask.astype(np.uint8)\n\nmask = probs >= THRESHOLD\nmask = postprocess(mask.astype(np.uint8))\n\n```\n\nI think the last cell will be helpful as a reference.\nhttps://www.kaggle.com/code/sugupoko/v30-4-group-ensemble-blend-blur-bright-lowco?scriptVersionId=300348690&cellId=16", "votes": null }, { "id": "3415453", "postDate": "02/28/2026 23:49:18", "content": "

Wow, awesome. This is a great post process. I see that you update each pixel prediction based on its 27 neighbors' median.

", "rawMarkdown": "Wow, awesome. This is a great post process. I see that you update each pixel prediction based on its 27 neighbors' median.", "votes": null }, { "id": "3415532", "postDate": "03/01/2026 01:18:27", "content": "

Hi @giorgioangelotti san

\n

I had forgotten about my experiment, I did analyze it.\nI wrote it up here — thank you!

\n

https://www.kaggle.com/competitions/vesuvius-challenge-surface-detection/discussion/679379

", "rawMarkdown": "Hi @giorgioangelotti san\n\nI had forgotten about my experiment, I did analyze it.\nI wrote it up here — thank you!\n\nhttps://www.kaggle.com/competitions/vesuvius-challenge-surface-detection/discussion/679379", "votes": null }, { "id": "3415705", "postDate": "03/01/2026 07:05:53", "content": "

Bravo for 18th place @sugupoko

", "rawMarkdown": "Bravo for 18th place @sugupoko", "votes": null } ], "comments": [ { "id": 3414940, "author_name": "sugupoko", "author_url": "", "post_date": "02/28/2026 01:06:27", "content": "

6 models. runtime 6h. single thread

", "votes": null, "replies": [ { "id": 3415362, "author_name": "rajeshthevar", "author_url": "", "post_date": "02/28/2026 19:01:10", "content": "

How? :O\nChecking your shared notebook.

", "votes": null, "replies": [ { "id": 3415441, "author_name": "sugupoko", "author_url": "", "post_date": "02/28/2026 23:25:08", "content": "

Z compression is very fast. In most cases, Z is redundant.

", "votes": null, "replies": [] } ] } ] }, { "id": 3414975, "author_name": "sugupoko", "author_url": "", "post_date": "02/28/2026 02:43:39", "content": "

code \nhttps://www.kaggle.com/code/sugupoko/v30-4-group-ensemble-blend-blur-bright-lowco?scriptVersionId=300017285

", "votes": null, "replies": [] }, { "id": 3415131, "author_name": "giorgioangelotti", "author_url": "", "post_date": "02/28/2026 08:15:09", "content": "

Interesting solution!

\n

Did you compress in Z every input sample, or you decided the axis of compression depending on some analysis?

", "votes": null, "replies": [ { "id": 3415148, "author_name": "sugupoko", "author_url": "", "post_date": "02/28/2026 09:16:15", "content": "

Thank you for your comment!!

\n

No, I didn’t choose it based on any analysis. Given the VRAM constraints of the RTX 5090, I just fixed it at around that range.

", "votes": null, "replies": [] }, { "id": 3415532, "author_name": "sugupoko", "author_url": "", "post_date": "03/01/2026 01:18:27", "content": "

Hi @giorgioangelotti san

\n

I had forgotten about my experiment, I did analyze it.\nI wrote it up here — thank you!

\n

https://www.kaggle.com/competitions/vesuvius-challenge-surface-detection/discussion/679379

", "votes": null, "replies": [] } ] }, { "id": 3415146, "author_name": "sugupoko", "author_url": "", "post_date": "02/28/2026 09:13:02", "content": "

Architecture : host model\nconfig :

\n", "votes": null, "replies": [] }, { "id": 3415444, "author_name": "cdeotte", "author_url": "", "post_date": "02/28/2026 23:28:41", "content": "

Congratulations on strong 18th place finish. Can you explain the \"median filter x7 times\"? I'm curious how this works. Do you ensemble models and/or use TTA and then take the median?

", "votes": null, "replies": [ { "id": 3415448, "author_name": "sugupoko", "author_url": "", "post_date": "02/28/2026 23:40:41", "content": "

Thank you for your comment!!!

\n

After ensemble and threshold-based binarization, I simply apply the following code.\nThe post-processing is minimal but very strong.

\n
from scipy.ndimage import median_filter\n\nPP_MEDIAN_ITER = 7\nPP_MEDIAN_SIZE = 3\n\ndef postprocess(mask: np.ndarray) -> np.ndarray:\n    mask = mask.copy()\n    for _ in range(PP_MEDIAN_ITER):\n        mask = (median_filter(mask, size=PP_MEDIAN_SIZE) > 0).astype(np.uint8)\n    return mask.astype(np.uint8)\n\nmask = probs >= THRESHOLD\nmask = postprocess(mask.astype(np.uint8))\n
\n

I think the last cell will be helpful as a reference.\nhttps://www.kaggle.com/code/sugupoko/v30-4-group-ensemble-blend-blur-bright-lowco?scriptVersionId=300348690&cellId=16

", "votes": null, "replies": [ { "id": 3415453, "author_name": "cdeotte", "author_url": "", "post_date": "02/28/2026 23:49:18", "content": "

Wow, awesome. This is a great post process. I see that you update each pixel prediction based on its 27 neighbors' median.

", "votes": null, "replies": [] } ] } ] }, { "id": 3415705, "author_name": "navneetbende", "author_url": "", "post_date": "03/01/2026 07:05:53", "content": "

Bravo for 18th place @sugupoko

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3414934": "Great competition—thanks to the Kaggle team for organizing and running everything so smoothly!", "3414940": "6 models. runtime 6h. single thread", "3414975": "code \nhttps://www.kaggle.com/code/sugupoko/v30-4-group-ensemble-blend-blur-bright-lowco?scriptVersionId=300017285", "3415131": "Interesting solution!\n\nDid you compress in Z every input sample, or you decided the axis of compression depending on some analysis?", "3415146": "Architecture : host model\nconfig : \n- Pretrained model: 200 epochs, trained on ~700 pseudo-labeled samples + 400 external samples.\n- Global-coverage model: 500 epochs, trained without pseudo-labeled data.\n- Specialized model: 100 epochs, trained on selected data (~100–200 samples).", "3415148": "Thank you for your comment!!\n\nNo, I didn’t choose it based on any analysis. Given the VRAM constraints of the RTX 5090, I just fixed it at around that range.", "3415362": "How? :O\nChecking your shared notebook.", "3415441": "Z compression is very fast. In most cases, Z is redundant.", "3415444": "Congratulations on strong 18th place finish. Can you explain the \"median filter x7 times\"? I'm curious how this works. Do you ensemble models and/or use TTA and then take the median?", "3415448": "Thank you for your comment!!!\n\nAfter ensemble and threshold-based binarization, I simply apply the following code.\nThe post-processing is minimal but very strong.\n\n```python\nfrom scipy.ndimage import median_filter\n\nPP_MEDIAN_ITER = 7\nPP_MEDIAN_SIZE = 3\n\ndef postprocess(mask: np.ndarray) -> np.ndarray:\n mask = mask.copy()\n for _ in range(PP_MEDIAN_ITER):\n mask = (median_filter(mask, size=PP_MEDIAN_SIZE) > 0).astype(np.uint8)\n return mask.astype(np.uint8)\n\nmask = probs >= THRESHOLD\nmask = postprocess(mask.astype(np.uint8))\n\n```\n\nI think the last cell will be helpful as a reference.\nhttps://www.kaggle.com/code/sugupoko/v30-4-group-ensemble-blend-blur-bright-lowco?scriptVersionId=300348690&cellId=16", "3415453": "Wow, awesome. This is a great post process. I see that you update each pixel prediction based on its 27 neighbors' median.", "3415532": "Hi @giorgioangelotti san\n\nI had forgotten about my experiment, I did analyze it.\nI wrote it up here — thank you!\n\nhttps://www.kaggle.com/competitions/vesuvius-challenge-surface-detection/discussion/679379", "3415705": "Bravo for 18th place @sugupoko" }, "source": "meta" }