{ "id": 414035, "title": "Has anyone experimented with techniques that can reduce edge effects?", "url": "/competitions/vesuvius-challenge-ink-detection/discussion/414035", "author_name": "", "post_date": "2023-05-31T05:28:18.467289900Z", "votes": 5, "comment_count": 3, "views": 0, "content": "

Please share any insights you have :)
\nI'll start by sharing this paper
\nand this repo

", "messages": [ { "id": "2281745", "postDate": "05/31/2023 05:28:18", "content": "

Please share any insights you have :)
\nI'll start by sharing this paper
\nand this repo

", "rawMarkdown": "Please share any insights you have :)\nI'll start by sharing this [paper](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0229839)\nand this [repo](https://github.com/Vooban/Smoothly-Blend-Image-Patches)", "votes": null }, { "id": "2281857", "postDate": "05/31/2023 07:16:33", "content": "

I think morphological transformations like opening and closing can be useful.
\nOpenCV Morphological transformations

", "rawMarkdown": "I think morphological transformations like opening and closing can be useful.\n[OpenCV Morphological transformations](https://docs.opencv.org/4.x/d9/d61/tutorial_py_morphological_ops.html)", "votes": null }, { "id": "2283495", "postDate": "06/01/2023 09:41:57", "content": "

Haven't investigated thoroughly, but isn't this:

\n
def make_infer_mask():\n    s = CFG.crop_size\n    f = CFG.crop_fade\n    x = np.linspace(-1, 1, s)\n    y = np.linspace(-1, 1, s)\n    xx, yy = np.meshgrid(x, y)\n    d = 1 - np.maximum(np.abs(xx), np.abs(yy))\n    d1 = np.clip(d, 0, f / s * 2)\n    d1 = d1 / d1.max()\n    infer_mask = d1\n    return infer_mask\n
\n

in combination with the code a bit further down probability[y0:y0 + crop_size, x0:x0 + crop_size] += k[b,0]*infer_mask doing that?

\n

@hengck23 notebook https://www.kaggle.com/code/hengck23/lb0-68-one-fold-stacked-unet/notebook?

", "rawMarkdown": "Haven't investigated thoroughly, but isn't this:\n\n```\ndef make_infer_mask():\n\ts = CFG.crop_size\n\tf = CFG.crop_fade\n\tx = np.linspace(-1, 1, s)\n\ty = np.linspace(-1, 1, s)\n\txx, yy = np.meshgrid(x, y)\n\td = 1 - np.maximum(np.abs(xx), np.abs(yy))\n\td1 = np.clip(d, 0, f / s * 2)\n\td1 = d1 / d1.max()\n\tinfer_mask = d1\n\treturn infer_mask\n```\n\nin combination with the code a bit further down `probability[y0:y0 + crop_size, x0:x0 + crop_size] += k[b,0]*infer_mask` doing that?\n\n@hengck23 notebook https://www.kaggle.com/code/hengck23/lb0-68-one-fold-stacked-unet/notebook?", "votes": null }, { "id": "2294930", "postDate": "06/10/2023 13:17:27", "content": "

I used the morphological transformation to reduce mask near edge, but it didn’t result in higher CV in my case.

", "rawMarkdown": "I used the morphological transformation to reduce mask near edge, but it didn’t result in higher CV in my case.", "votes": null } ], "comments": [ { "id": 2281857, "author_name": "vijaybj", "author_url": "", "post_date": "05/31/2023 07:16:33", "content": "

I think morphological transformations like opening and closing can be useful.
\nOpenCV Morphological transformations

", "votes": null, "replies": [ { "id": 2294930, "author_name": "clearwaterkzk", "author_url": "", "post_date": "06/10/2023 13:17:27", "content": "

I used the morphological transformation to reduce mask near edge, but it didn’t result in higher CV in my case.

", "votes": null, "replies": [] } ] }, { "id": 2283495, "author_name": "lucasvw", "author_url": "", "post_date": "06/01/2023 09:41:57", "content": "

Haven't investigated thoroughly, but isn't this:

\n
def make_infer_mask():\n    s = CFG.crop_size\n    f = CFG.crop_fade\n    x = np.linspace(-1, 1, s)\n    y = np.linspace(-1, 1, s)\n    xx, yy = np.meshgrid(x, y)\n    d = 1 - np.maximum(np.abs(xx), np.abs(yy))\n    d1 = np.clip(d, 0, f / s * 2)\n    d1 = d1 / d1.max()\n    infer_mask = d1\n    return infer_mask\n
\n

in combination with the code a bit further down probability[y0:y0 + crop_size, x0:x0 + crop_size] += k[b,0]*infer_mask doing that?

\n

@hengck23 notebook https://www.kaggle.com/code/hengck23/lb0-68-one-fold-stacked-unet/notebook?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2281745": "Please share any insights you have :)\nI'll start by sharing this [paper](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0229839)\nand this [repo](https://github.com/Vooban/Smoothly-Blend-Image-Patches)", "2281857": "I think morphological transformations like opening and closing can be useful.\n[OpenCV Morphological transformations](https://docs.opencv.org/4.x/d9/d61/tutorial_py_morphological_ops.html)", "2283495": "Haven't investigated thoroughly, but isn't this:\n\n```\ndef make_infer_mask():\n\ts = CFG.crop_size\n\tf = CFG.crop_fade\n\tx = np.linspace(-1, 1, s)\n\ty = np.linspace(-1, 1, s)\n\txx, yy = np.meshgrid(x, y)\n\td = 1 - np.maximum(np.abs(xx), np.abs(yy))\n\td1 = np.clip(d, 0, f / s * 2)\n\td1 = d1 / d1.max()\n\tinfer_mask = d1\n\treturn infer_mask\n```\n\nin combination with the code a bit further down `probability[y0:y0 + crop_size, x0:x0 + crop_size] += k[b,0]*infer_mask` doing that?\n\n@hengck23 notebook https://www.kaggle.com/code/hengck23/lb0-68-one-fold-stacked-unet/notebook?", "2294930": "I used the morphological transformation to reduce mask near edge, but it didn’t result in higher CV in my case." }, "source": "meta" }