{ "id": 424900, "title": "Strategies to ensemble multiple masks?", "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/424900", "author_name": "Yassine Alouini", "post_date": "2023-07-16T08:41:11.356000", "votes": 1, "comment_count": 8, "views": 0, "content": "

So far, the strategy I thought about is the following:

\n\n

Another strategy is to take the average pixels of the masks of course.

\n

What other strategies are you using?

\n

As suggested by @atamazian, weighted masks fusion (WMF) can be used. It is an adaptation to masks of the weighted boxes fusion method (WBF in short) that you can read more about here.

\n

An implementation can be found here.

", "messages": [ { "id": 2351464, "postDate": "2023-07-20T06:52:57.373Z", "content": "

You either match multiple masks of multiple models by IoU or you can stack them, create semantic segmentation mask, blend them and separate them once again.

", "rawMarkdown": "You either match multiple masks of multiple models by IoU or you can stack them, create semantic segmentation mask, blend them and separate them once again.", "votes": 1, "replies": [ { "id": 2351828, "postDate": "2023-07-20T12:06:16.290Z", "content": "

Blending and separating is a great idea, thanks for the suggestion. What types of algorithm you use for the separation step ?

", "rawMarkdown": "Blending and separating is a great idea, thanks for the suggestion. What types of algorithm you use for the separation step ? " } ] }, { "id": 2346928, "postDate": "2023-07-16T16:23:51.700Z", "content": "

This notebook has a good package for an ensembling: https://www.kaggle.com/code/markunys/8th-place-solution-inference

", "rawMarkdown": "This notebook has a good package for an ensembling: https://www.kaggle.com/code/markunys/8th-place-solution-inference", "votes": 1, "replies": [ { "id": 2347975, "postDate": "2023-07-17T10:19:23.740Z", "content": "

Will have a look, thanks @atamazian!

", "rawMarkdown": "Will have a look, thanks @atamazian!", "replies": [ { "id": 2348025, "postDate": "2023-07-17T10:51:50.583Z", "content": "

Indeed, the following dataset contains WMF (weighted masks fusion) https://www.kaggle.com/datasets/markunys/ensemble-boxes

", "rawMarkdown": "Indeed, the following dataset contains WMF (weighted masks fusion) https://www.kaggle.com/datasets/markunys/ensemble-boxes" } ] } ] }, { "id": 2346435, "postDate": "2023-07-16T08:41:11.357Z", "content": "

So far, the strategy I thought about is the following:

\n\n

Another strategy is to take the average pixels of the masks of course.

\n

What other strategies are you using?

\n

As suggested by @atamazian, weighted masks fusion (WMF) can be used. It is an adaptation to masks of the weighted boxes fusion method (WBF in short) that you can read more about here.

\n

An implementation can be found here.

", "rawMarkdown": "So far, the strategy I thought about is the following:\n\n- use mask 1 if score 1 is better\n- otherwise use mask 2\n\nAnother strategy is to take the average pixels of the masks of course.\n\nWhat other strategies are you using?\n\n\nAs suggested by @atamazian, **weighted masks fusion** (WMF) can be used. It is an adaptation to masks of the **weighted boxes fusion** method (WBF in short) that you can read more about [here](https://arxiv.org/pdf/1910.13302.pdf).\n\nAn implementation can be found [here](https://www.kaggle.com/code/mistag/sartorius-tta-with-weighted-segments-fusion).", "votes": 1 }, { "id": 2346478, "postDate": "2023-07-16T09:27:42.233Z", "content": "

Not directly related but this notebook explains how TTA works for segmentation (should be applied twice so that the predicted mask is in the correct position before averaging) => https://www.kaggle.com/code/joshi98kishan/let-s-understand-tta-in-segmentation

", "rawMarkdown": "Not directly related but this notebook explains how TTA works for segmentation (should be applied twice so that the predicted mask is in the correct position before averaging) => https://www.kaggle.com/code/joshi98kishan/let-s-understand-tta-in-segmentation" }, { "id": 2355208, "postDate": "2023-07-23T07:12:31.503Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 2355210, "postDate": "2023-07-23T07:17:37.093Z", "content": "

My code based on the notebook mentioned above https://www.kaggle.com/datasets/markunys/ensemble-boxes and I changed some parts to solve the \"out of memory\" problem. The original code save all the mask during the process costing too much memory.

", "rawMarkdown": "My code based on the notebook mentioned above [https://www.kaggle.com/datasets/markunys/ensemble-boxes](url) and I changed some parts to solve the \"out of memory\" problem. The original code save all the mask during the process costing too much memory." } ] } ], "comments": [ { "id": 2351464, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-07-20T06:52:57.373000", "content": "

You either match multiple masks of multiple models by IoU or you can stack them, create semantic segmentation mask, blend them and separate them once again.

", "votes": 1, "replies": [ { "id": 2351828, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2023-07-20T12:06:16.290000", "content": "

Blending and separating is a great idea, thanks for the suggestion. What types of algorithm you use for the separation step ?

", "votes": 0, "replies": [] } ] }, { "id": 2346928, "author_name": "Araik Tamazian", "author_url": "", "post_date": "2023-07-16T16:23:51.700000", "content": "

This notebook has a good package for an ensembling: https://www.kaggle.com/code/markunys/8th-place-solution-inference

", "votes": 1, "replies": [ { "id": 2347975, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2023-07-17T10:19:23.740000", "content": "

Will have a look, thanks @atamazian!

", "votes": 0, "replies": [ { "id": 2348025, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2023-07-17T10:51:50.583000", "content": "

Indeed, the following dataset contains WMF (weighted masks fusion) https://www.kaggle.com/datasets/markunys/ensemble-boxes

", "votes": 0, "replies": [] } ] } ] }, { "id": 2346478, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2023-07-16T09:27:42.233000", "content": "

Not directly related but this notebook explains how TTA works for segmentation (should be applied twice so that the predicted mask is in the correct position before averaging) => https://www.kaggle.com/code/joshi98kishan/let-s-understand-tta-in-segmentation

", "votes": 0, "replies": [] }, { "id": 2355208, "author_name": "", "author_url": "", "post_date": "2023-07-23T07:12:31.503000", "content": "", "votes": 0, "replies": [ { "id": 2355210, "author_name": "Blue He", "author_url": "", "post_date": "2023-07-23T07:17:37.093000", "content": "

My code based on the notebook mentioned above https://www.kaggle.com/datasets/markunys/ensemble-boxes and I changed some parts to solve the \"out of memory\" problem. The original code save all the mask during the process costing too much memory.

", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2351464": "You either match multiple masks of multiple models by IoU or you can stack them, create semantic segmentation mask, blend them and separate them once again.", "2346928": "This notebook has a good package for an ensembling: https://www.kaggle.com/code/markunys/8th-place-solution-inference", "2346435": "So far, the strategy I thought about is the following:\n\n- use mask 1 if score 1 is better\n- otherwise use mask 2\n\nAnother strategy is to take the average pixels of the masks of course.\n\nWhat other strategies are you using?\n\n\nAs suggested by @atamazian, **weighted masks fusion** (WMF) can be used. It is an adaptation to masks of the **weighted boxes fusion** method (WBF in short) that you can read more about [here](https://arxiv.org/pdf/1910.13302.pdf).\n\nAn implementation can be found [here](https://www.kaggle.com/code/mistag/sartorius-tta-with-weighted-segments-fusion).", "2346478": "Not directly related but this notebook explains how TTA works for segmentation (should be applied twice so that the predicted mask is in the correct position before averaging) => https://www.kaggle.com/code/joshi98kishan/let-s-understand-tta-in-segmentation", "2355208": "" } }