{ "id": 312077, "title": "Best practices to handle different image sizes", "url": "/competitions/happy-whale-and-dolphin/discussion/312077", "author_name": "", "post_date": "2022-03-10T08:42:46.314652500Z", "votes": 6, "comment_count": 6, "views": 0, "content": "

Which best practices could you advice in order to handle different image resolution, ranging from 68x75 to 5399x3599 pixels ?

", "messages": [ { "id": "1717821", "postDate": "03/10/2022 08:42:46", "content": "

Which best practices could you advice in order to handle different image resolution, ranging from 68x75 to 5399x3599 pixels ?

", "rawMarkdown": "Which best practices could you advice in order to handle different image resolution, ranging from 68x75 to 5399x3599 pixels ?", "votes": null }, { "id": "1717861", "postDate": "03/10/2022 09:35:23", "content": "

Resize it to one size :)

", "rawMarkdown": "Resize it to one size :)", "votes": null }, { "id": "1718007", "postDate": "03/10/2022 12:33:47", "content": "

I agree. It is a solution.
\nBut do you really think that is a best practice.

", "rawMarkdown": "I agree. It is a solution.\nBut do you really think that is a best practice.", "votes": null }, { "id": "1718017", "postDate": "03/10/2022 12:43:52", "content": "

I'm afraid yes, I didn't have any success stories with multiresolution classification, but It works for detectors

", "rawMarkdown": "I'm afraid yes, I didn't have any success stories with multiresolution classification, but It works for detectors", "votes": null }, { "id": "1718097", "postDate": "03/10/2022 14:02:16", "content": "

Multiresolution classification works in very specific scenerio, it did work in a few top solutions in 2020 and 2021 Google Landmark Recognition/Identification

", "rawMarkdown": "Multiresolution classification works in very specific scenerio, it did work in a few top solutions in 2020 and 2021 Google Landmark Recognition/Identification", "votes": null }, { "id": "1718102", "postDate": "03/10/2022 14:06:24", "content": "

I would say that resize is a very strong thing, which is incredibly underestimated with its true potential. For example, the biggest factor is interpolation, you would not want a linear interpolation on a human video upscaled, using cubic interpolation would look much much better. Another factor is the aspect ratio, in some cases where the object are very much differentiated by their size and not by its visual features (visual features like texture, smoothness, color), preserving the aspect ratio would be one of the only differentiated feature model will be learning from.

", "rawMarkdown": "I would say that resize is a very strong thing, which is incredibly underestimated with its true potential. For example, the biggest factor is interpolation, you would not want a linear interpolation on a human video upscaled, using cubic interpolation would look much much better. Another factor is the aspect ratio, in some cases where the object are very much differentiated by their size and not by its visual features (visual features like texture, smoothness, color), preserving the aspect ratio would be one of the only differentiated feature model will be learning from.", "votes": null }, { "id": "1725806", "postDate": "03/17/2022 12:55:47", "content": "

This question is connected to the HappyWhale Competition. Because pictures represent the salient items at many scales (very closed, or far away), I have decided to manually label the pictures. I publish a dataset of the pictures sampling strategy.

", "rawMarkdown": "This question is connected to the [HappyWhale Competition](https://www.kaggle.com/c/happy-whale-and-dolphin). Because pictures represent the salient items at many scales (very closed, or far away), I have decided to manually label the pictures. I publish a dataset of the [pictures sampling strategy](https://www.kaggle.com/datasets/chasset/pictures-sampling-by-specy-and-individual).", "votes": null } ], "comments": [ { "id": 1717861, "author_name": "kwentar", "author_url": "", "post_date": "03/10/2022 09:35:23", "content": "

Resize it to one size :)

", "votes": null, "replies": [ { "id": 1718007, "author_name": "chasset", "author_url": "", "post_date": "03/10/2022 12:33:47", "content": "

I agree. It is a solution.
\nBut do you really think that is a best practice.

", "votes": null, "replies": [] }, { "id": 1718017, "author_name": "kwentar", "author_url": "", "post_date": "03/10/2022 12:43:52", "content": "

I'm afraid yes, I didn't have any success stories with multiresolution classification, but It works for detectors

", "votes": null, "replies": [] }, { "id": 1718097, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/10/2022 14:02:16", "content": "

Multiresolution classification works in very specific scenerio, it did work in a few top solutions in 2020 and 2021 Google Landmark Recognition/Identification

", "votes": null, "replies": [] } ] }, { "id": 1718102, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/10/2022 14:06:24", "content": "

I would say that resize is a very strong thing, which is incredibly underestimated with its true potential. For example, the biggest factor is interpolation, you would not want a linear interpolation on a human video upscaled, using cubic interpolation would look much much better. Another factor is the aspect ratio, in some cases where the object are very much differentiated by their size and not by its visual features (visual features like texture, smoothness, color), preserving the aspect ratio would be one of the only differentiated feature model will be learning from.

", "votes": null, "replies": [] }, { "id": 1725806, "author_name": "chasset", "author_url": "", "post_date": "03/17/2022 12:55:47", "content": "

This question is connected to the HappyWhale Competition. Because pictures represent the salient items at many scales (very closed, or far away), I have decided to manually label the pictures. I publish a dataset of the pictures sampling strategy.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1717821": "Which best practices could you advice in order to handle different image resolution, ranging from 68x75 to 5399x3599 pixels ?", "1717861": "Resize it to one size :)", "1718007": "I agree. It is a solution.\nBut do you really think that is a best practice.", "1718017": "I'm afraid yes, I didn't have any success stories with multiresolution classification, but It works for detectors", "1718097": "Multiresolution classification works in very specific scenerio, it did work in a few top solutions in 2020 and 2021 Google Landmark Recognition/Identification", "1718102": "I would say that resize is a very strong thing, which is incredibly underestimated with its true potential. For example, the biggest factor is interpolation, you would not want a linear interpolation on a human video upscaled, using cubic interpolation would look much much better. Another factor is the aspect ratio, in some cases where the object are very much differentiated by their size and not by its visual features (visual features like texture, smoothness, color), preserving the aspect ratio would be one of the only differentiated feature model will be learning from.", "1725806": "This question is connected to the [HappyWhale Competition](https://www.kaggle.com/c/happy-whale-and-dolphin). Because pictures represent the salient items at many scales (very closed, or far away), I have decided to manually label the pictures. I publish a dataset of the [pictures sampling strategy](https://www.kaggle.com/datasets/chasset/pictures-sampling-by-specy-and-individual)." }, "source": "meta" }