{ "id": 311990, "title": "Annotate Dataset ", "url": "/competitions/ultra-mnist/discussion/311990", "author_name": "", "post_date": "2022-03-09T21:10:32.534037100Z", "votes": 2, "comment_count": 6, "views": 0, "content": "

Not sure how good this idea is, but you can manually crop a few samples of each digit from the images given in the current dataset.

\n

\"Screenshot \"Screenshot

\n

Now you can try generating the similar data on your own:

\n\n
Generated data sample:
\n

\"image\"

\n

Digit: 7
\nBounding Box: 0, 0, 20, 20

\n

I hope generating dataset this way on your own is not counted towards using external data (as we are just manipulating the given dataset)

", "messages": [ { "id": "1717358", "postDate": "03/09/2022 21:10:32", "content": "

Not sure how good this idea is, but you can manually crop a few samples of each digit from the images given in the current dataset.

\n

\"Screenshot \"Screenshot

\n

Now you can try generating the similar data on your own:

\n\n
Generated data sample:
\n

\"image\"

\n

Digit: 7
\nBounding Box: 0, 0, 20, 20

\n

I hope generating dataset this way on your own is not counted towards using external data (as we are just manipulating the given dataset)

", "rawMarkdown": "Not sure how good this idea is, but you can manually crop a few samples of each digit from the images given in the current dataset.\n\n\"Screenshot \"Screenshot\n\n\nNow you can try generating the similar data on your own:\n- generate a black image (use tools like openCV), and randomly select the bounding box coordinate regions (say, x0, y0, x1, y1 = 0, 0, 20, 20), and randomly select a digit (say 7)\n- Augment the crop sample of the digit (here 7) and paste it on the black image in the bounding box region (here 0,0,20,20)\n- You have bounding box and ground truth of the new generated image.\n- Repeat the steps a couple of times to have multiple digits on single image (keeping sum of digits on the image in valid range)\n- Fine-tune your regressor, and image classifier on the newly generated data.\n\n\n\n##### Generated data sample:\n\"image\"\n\nDigit: 7\nBounding Box: 0, 0, 20, 20\n\n\n\nI hope generating dataset this way on your own is not counted towards using external data (as we are just manipulating the given dataset)", "votes": null }, { "id": "1717366", "postDate": "03/09/2022 21:23:07", "content": "

As far as I understand you can generate NEW dataset using competition dataset. So you can transform it as you can … and there is no limitations in this area. You cannot use external data (pixels which comes from different dataset).

", "rawMarkdown": "As far as I understand you can generate NEW dataset using competition dataset. So you can transform it as you can ... and there is no limitations in this area. You cannot use external data (pixels which comes from different dataset).", "votes": null }, { "id": "1717367", "postDate": "03/09/2022 21:31:42", "content": "

Agreed. And I am using only the given dataset to create new data with the corresponding labels.

", "rawMarkdown": "Agreed. And I am using only the given dataset to create new data with the corresponding labels.", "votes": null }, { "id": "1717385", "postDate": "03/09/2022 22:17:08", "content": "

This is a good idea and Data Augmentation/Generation should not lead to any issue as the data generated is from the same dataset.

", "rawMarkdown": "This is a good idea and Data Augmentation/Generation should not lead to any issue as the data generated is from the same dataset.", "votes": null }, { "id": "1717390", "postDate": "03/09/2022 22:23:48", "content": "

In 20 minutes I will share notebook - keep it easy :) No model required …. you can use just old computer vision techniques to generate everything you need :)

", "rawMarkdown": "In 20 minutes I will share notebook - keep it easy :) No model required .... you can use just old computer vision techniques to generate everything you need :)", "votes": null }, { "id": "1717405", "postDate": "03/09/2022 22:43:51", "content": "

@remekkinas This is really helpful. Thanks a lot! :)

", "rawMarkdown": "remekkinas This is really helpful. Thanks a lot! :)", "votes": null }, { "id": "1717445", "postDate": "03/09/2022 23:55:12", "content": "

As I said … no model, annotating is required - just image manipulations: https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required

", "rawMarkdown": "As I said ... no model, annotating is required - just image manipulations: https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required", "votes": null } ], "comments": [ { "id": 1717366, "author_name": "remekkinas", "author_url": "", "post_date": "03/09/2022 21:23:07", "content": "

As far as I understand you can generate NEW dataset using competition dataset. So you can transform it as you can … and there is no limitations in this area. You cannot use external data (pixels which comes from different dataset).

", "votes": null, "replies": [ { "id": 1717367, "author_name": "harshraj22", "author_url": "", "post_date": "03/09/2022 21:31:42", "content": "

Agreed. And I am using only the given dataset to create new data with the corresponding labels.

", "votes": null, "replies": [] } ] }, { "id": 1717385, "author_name": "shivkumarganesh", "author_url": "", "post_date": "03/09/2022 22:17:08", "content": "

This is a good idea and Data Augmentation/Generation should not lead to any issue as the data generated is from the same dataset.

", "votes": null, "replies": [ { "id": 1717390, "author_name": "remekkinas", "author_url": "", "post_date": "03/09/2022 22:23:48", "content": "

In 20 minutes I will share notebook - keep it easy :) No model required …. you can use just old computer vision techniques to generate everything you need :)

", "votes": null, "replies": [] }, { "id": 1717405, "author_name": "shivkumarganesh", "author_url": "", "post_date": "03/09/2022 22:43:51", "content": "

@remekkinas This is really helpful. Thanks a lot! :)

", "votes": null, "replies": [] }, { "id": 1717445, "author_name": "remekkinas", "author_url": "", "post_date": "03/09/2022 23:55:12", "content": "

As I said … no model, annotating is required - just image manipulations: https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1717358": "Not sure how good this idea is, but you can manually crop a few samples of each digit from the images given in the current dataset.\n\n\"Screenshot \"Screenshot\n\n\nNow you can try generating the similar data on your own:\n- generate a black image (use tools like openCV), and randomly select the bounding box coordinate regions (say, x0, y0, x1, y1 = 0, 0, 20, 20), and randomly select a digit (say 7)\n- Augment the crop sample of the digit (here 7) and paste it on the black image in the bounding box region (here 0,0,20,20)\n- You have bounding box and ground truth of the new generated image.\n- Repeat the steps a couple of times to have multiple digits on single image (keeping sum of digits on the image in valid range)\n- Fine-tune your regressor, and image classifier on the newly generated data.\n\n\n\n##### Generated data sample:\n\"image\"\n\nDigit: 7\nBounding Box: 0, 0, 20, 20\n\n\n\nI hope generating dataset this way on your own is not counted towards using external data (as we are just manipulating the given dataset)", "1717366": "As far as I understand you can generate NEW dataset using competition dataset. So you can transform it as you can ... and there is no limitations in this area. You cannot use external data (pixels which comes from different dataset).", "1717367": "Agreed. And I am using only the given dataset to create new data with the corresponding labels.", "1717385": "This is a good idea and Data Augmentation/Generation should not lead to any issue as the data generated is from the same dataset.", "1717390": "In 20 minutes I will share notebook - keep it easy :) No model required .... you can use just old computer vision techniques to generate everything you need :)", "1717405": "remekkinas This is really helpful. Thanks a lot! :)", "1717445": "As I said ... no model, annotating is required - just image manipulations: https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required" }, "source": "meta" }