Hi everyone!
\nSince the data is being reset and everyone is sharing their approach to clean the background, sharing mine here. I use a simple block matching technique to check a patch with a pure black/white patch and decide to invert colours or not based on the RMSE value.
Hi everyone!
\nSince the data is being reset and everyone is sharing their approach to clean the background, sharing mine here. I use a simple block matching technique to check a patch with a pure black/white patch and decide to invert colours or not based on the RMSE value.
Interesting Idea.
\nUnfortunately, it does suffer from the same problem as the \"Ultra-mnist black\" method
\nids = [\n\"bfnwpcvgdt\",\n\"vhmqewzxef\",\n\"wuumlogxge\",\n\"gcpxitjbvc\",\n\"zwhazcdilo\",\n\"zafyjsvftb\",\n\"exuzvpmqfe\",\n\"fsnaafcqor\",\n]\nfor id in ids:\n pth = f'../input/ultramnist2048x2048/train/{id}.jpeg'\n im_cv = cv2.cvtColor(cv2.imread(pth), cv2.COLOR_BGR2RGB)\n im_orig = np.asarray(im_cv)\n im = clean_bg(pth, 4) # 51 ms ± 5.35 ms per loop\n fig = plt.figure(figsize=(10,10))\n plt.subplot(2, 2, 1)\n plt.imshow(im_orig, cmap='Greys')\n plt.title('Before')\n plt.subplot(2, 2, 2)\n plt.imshow(im, cmap='Greys')\n plt.title('After')\n display(fig)\n plt.close()\nWow, nice catch! With clean_bg(pth, 128) and it behaves like a Canny edge filter for the listed images. Maybe with an extra Contour detection+fill step, it should still work.
When I try to use it on the current (new) dataset, I get this:
\nBxViolation: basebx of pybx.basics: Expected single element in coords, got [array([[ 0., 0., 128., 128.],\n [128., 0., 256., 128.],\n [256., 0., 384., 128.],\n...\n
I converted cv2.COLOR_BGR2RGB to cv2.COLOR_BGR2GRAY but that didn't work as well.
\nWhat might be wrong?
", "rawMarkdown": "When I try to use it on the current (new) dataset, I get this:\n\n`BxViolation: basebx of pybx.basics: Expected single element in coords, got [array([[ 0., 0., 128., 128.],\n [128., 0., 256., 128.],\n [256., 0., 384., 128.],\n...\n`\n\nI converted cv2.COLOR_BGR2RGB to cv2.COLOR_BGR2GRAY but that didn't work as well.\n\nWhat might be wrong?", "votes": null }, { "id": "1753440", "postDate": "04/12/2022 20:33:33", "content": "Looks like a change in the API caused this. You can check the updated notebook to see how it works with the new dataset.
", "rawMarkdown": "Looks like a change in the API caused this. You can check the [updated notebook](https://www.kaggle.com/code/thatgeeman/block-matching-to-clean-ultramnist-background) to see how it works with the new dataset.", "votes": null } ], "comments": [ { "id": 1723412, "author_name": "towzeur", "author_url": "", "post_date": "03/15/2022 11:54:58", "content": "Interesting Idea.
\nUnfortunately, it does suffer from the same problem as the \"Ultra-mnist black\" method
\nids = [\n\"bfnwpcvgdt\",\n\"vhmqewzxef\",\n\"wuumlogxge\",\n\"gcpxitjbvc\",\n\"zwhazcdilo\",\n\"zafyjsvftb\",\n\"exuzvpmqfe\",\n\"fsnaafcqor\",\n]\nfor id in ids:\n pth = f'../input/ultramnist2048x2048/train/{id}.jpeg'\n im_cv = cv2.cvtColor(cv2.imread(pth), cv2.COLOR_BGR2RGB)\n im_orig = np.asarray(im_cv)\n im = clean_bg(pth, 4) # 51 ms ± 5.35 ms per loop\n fig = plt.figure(figsize=(10,10))\n plt.subplot(2, 2, 1)\n plt.imshow(im_orig, cmap='Greys')\n plt.title('Before')\n plt.subplot(2, 2, 2)\n plt.imshow(im, cmap='Greys')\n plt.title('After')\n display(fig)\n plt.close()\nWow, nice catch! With clean_bg(pth, 128) and it behaves like a Canny edge filter for the listed images. Maybe with an extra Contour detection+fill step, it should still work.
When I try to use it on the current (new) dataset, I get this:
\nBxViolation: basebx of pybx.basics: Expected single element in coords, got [array([[ 0., 0., 128., 128.],\n [128., 0., 256., 128.],\n [256., 0., 384., 128.],\n...\n
I converted cv2.COLOR_BGR2RGB to cv2.COLOR_BGR2GRAY but that didn't work as well.
\nWhat might be wrong?
", "votes": null, "replies": [ { "id": 1753440, "author_name": "thatgeeman", "author_url": "", "post_date": "04/12/2022 20:33:33", "content": "Looks like a change in the API caused this. You can check the updated notebook to see how it works with the new dataset.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1723368": "Hi everyone! \nSince the data is being reset and everyone is sharing their approach to clean the background, sharing mine [here](https://www.kaggle.com/thatgeeman/block-matching-to-clean-ultramnist-background). I use a simple block matching technique to check a patch with a pure black/white patch and decide to invert colours or not based on the RMSE value.", "1723412": "Interesting Idea.\n\nUnfortunately, it does suffer from the same problem as the \"Ultra-mnist black\" method\n\n```python\nids = [\n\"bfnwpcvgdt\",\n\"vhmqewzxef\",\n\"wuumlogxge\",\n\"gcpxitjbvc\",\n\"zwhazcdilo\",\n\"zafyjsvftb\",\n\"exuzvpmqfe\",\n\"fsnaafcqor\",\n]\nfor id in ids:\n pth = f'../input/ultramnist2048x2048/train/{id}.jpeg'\n im_cv = cv2.cvtColor(cv2.imread(pth), cv2.COLOR_BGR2RGB)\n im_orig = np.asarray(im_cv)\n im = clean_bg(pth, 4) # 51 ms ± 5.35 ms per loop\n fig = plt.figure(figsize=(10,10))\n plt.subplot(2, 2, 1)\n plt.imshow(im_orig, cmap='Greys')\n plt.title('Before')\n plt.subplot(2, 2, 2)\n plt.imshow(im, cmap='Greys')\n plt.title('After')\n display(fig)\n plt.close()\n```", "1723466": "Wow, nice catch! With `clean_bg(pth, 128)` and it behaves like a Canny edge filter for the listed images. Maybe with an extra Contour detection+fill step, it should still work.", "1748021": "When I try to use it on the current (new) dataset, I get this:\n\n`BxViolation: basebx of pybx.basics: Expected single element in coords, got [array([[ 0., 0., 128., 128.],\n [128., 0., 256., 128.],\n [256., 0., 384., 128.],\n...\n`\n\nI converted cv2.COLOR_BGR2RGB to cv2.COLOR_BGR2GRAY but that didn't work as well.\n\nWhat might be wrong?", "1753440": "Looks like a change in the API caused this. You can check the [updated notebook](https://www.kaggle.com/code/thatgeeman/block-matching-to-clean-ultramnist-background) to see how it works with the new dataset." }, "source": "meta" }