In this competitions, most of the image size are 3000 3000, but same not. for training, the general method is to unify the image to 256256 or 512*512. But compared with original image, there are some information loss. Any better method such as Super resolution reconstruction can be applied or just unify the image to the same size?(As a green hand)
", "messages": [ { "id": "1874179", "postDate": "07/28/2022 05:44:39", "content": "In this competitions, most of the image size are 3000 3000, but same not. for training, the general method is to unify the image to 256256 or 512*512. But compared with original image, there are some information loss. Any better method such as Super resolution reconstruction can be applied or just unify the image to the same size?(As a green hand)
", "rawMarkdown": "In this competitions, most of the image size are 3000 *3000, but same not. for training, the general method is to unify the image to 256*256 or 512*512. But compared with original image, there are some information loss. Any better method such as Super resolution reconstruction can be applied or just unify the image to the same size?(As a green hand)", "votes": null }, { "id": "1874192", "postDate": "07/28/2022 06:02:26", "content": "In the testing dataset, there are also images that are 160x160 px: Quoting: \" The HuBMAP images range in size from 4500x4500 down to 160x160 pixels.\" What I usually do in such a case is huge pictures sliced on tiles that there is as close as it gets that every tile is 512x512 and then scale down by 1/2 to match 256x256. If there are some picture that is 160x160 I joust scale them up to match 256x256. If you do it like that you will not get \"TO MUCH\" of an information loss since the scale ratio is around [-0.5, +0.5] and you boost your training since you scale down huge pictures.
\nIMHO: Don't worry about Super-resolution reconstruction, since there is not much time left for submissions, I don't think you will benefit from that anyway.
", "rawMarkdown": "In the testing dataset, there are also images that are 160x160 px: Quoting: \" The HuBMAP images range in size from 4500x4500 down to 160x160 pixels.\" What I usually do in such a case is huge pictures sliced on tiles that there is as close as it gets that every tile is 512x512 and then scale down by 1/2 to match 256x256. If there are some picture that is 160x160 I joust scale them up to match 256x256. If you do it like that you will not get \"TO MUCH\" of an information loss since the scale ratio is around [-0.5, +0.5] and you boost your training since you scale down huge pictures.\n\nIMHO: Don't worry about Super-resolution reconstruction, since there is not much time left for submissions, I don't think you will benefit from that anyway.", "votes": null }, { "id": "1874947", "postDate": "07/28/2022 16:05:17", "content": "introduce the Transformer sequence feature extraction network and use encoding and decoding to realize that each pixel has the global receptive field.
", "rawMarkdown": "introduce the Transformer sequence feature extraction network and use encoding and decoding to realize that each pixel has the global receptive field.", "votes": null }, { "id": "1876733", "postDate": "07/30/2022 03:52:27", "content": "Thanks a lot, it is really helpful, I will try it.
", "rawMarkdown": "Thanks a lot, it is really helpful, I will try it.", "votes": null }, { "id": "1876734", "postDate": "07/30/2022 03:53:04", "content": "Thanks a lot, I will try the transformer model soon.
", "rawMarkdown": "Thanks a lot, I will try the transformer model soon.", "votes": null }, { "id": "1876784", "postDate": "07/30/2022 04:44:38", "content": "If you find comments helpful you can upvote them.
", "rawMarkdown": "If you find comments helpful you can upvote them.", "votes": null }, { "id": "1879296", "postDate": "08/01/2022 01:25:45", "content": "All hpa images are at 0.4 m pixels
\nduring testing you can scale the test images to -> test_image_height * test_pixel_size * ( model_input_shape / (3000 * 0.4) )
In the testing dataset, there are also images that are 160x160 px: Quoting: \" The HuBMAP images range in size from 4500x4500 down to 160x160 pixels.\" What I usually do in such a case is huge pictures sliced on tiles that there is as close as it gets that every tile is 512x512 and then scale down by 1/2 to match 256x256. If there are some picture that is 160x160 I joust scale them up to match 256x256. If you do it like that you will not get \"TO MUCH\" of an information loss since the scale ratio is around [-0.5, +0.5] and you boost your training since you scale down huge pictures.
\nIMHO: Don't worry about Super-resolution reconstruction, since there is not much time left for submissions, I don't think you will benefit from that anyway.
", "votes": null, "replies": [ { "id": 1876733, "author_name": "pengbing", "author_url": "", "post_date": "07/30/2022 03:52:27", "content": "Thanks a lot, it is really helpful, I will try it.
", "votes": null, "replies": [] }, { "id": 1876784, "author_name": "urosjarc", "author_url": "", "post_date": "07/30/2022 04:44:38", "content": "If you find comments helpful you can upvote them.
", "votes": null, "replies": [] } ] }, { "id": 1874947, "author_name": "corex0", "author_url": "", "post_date": "07/28/2022 16:05:17", "content": "introduce the Transformer sequence feature extraction network and use encoding and decoding to realize that each pixel has the global receptive field.
", "votes": null, "replies": [ { "id": 1876734, "author_name": "pengbing", "author_url": "", "post_date": "07/30/2022 03:53:04", "content": "Thanks a lot, I will try the transformer model soon.
", "votes": null, "replies": [] } ] }, { "id": 1879296, "author_name": "rarun2596", "author_url": "", "post_date": "08/01/2022 01:25:45", "content": "All hpa images are at 0.4 m pixels
\nduring testing you can scale the test images to -> test_image_height * test_pixel_size * ( model_input_shape / (3000 * 0.4) )