![\"Gpu\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F17954496%2Fb86cc4dc8f0e305c871413882cc9c773%2Fgpus.png?generation=1761677396359342&alt=media\")
Maybe it’s the data, maybe the metric.\nWhat do you think makes this challenge difficult. share your insights!
", "messages": [ { "id": 3310187, "postDate": "2025-11-02T10:12:00.240Z", "content": "One thing that I have found is that the authentic images should have empty masks, masks filled with 0 with the same size as the image. Without those masks you can't train a model because you need your model trained and tested with all the cases. You have to prepare those masks before using your models.
\nAnother point that makes this competition hard is that some segmentation models work with masks that have the same file extension that the images that are used as imput, so we have to invest time transforming from .npy to .png
", "rawMarkdown": "One thing that I have found is that the authentic images should have empty masks, masks filled with 0 with the same size as the image. Without those masks you can't train a model because you need your model trained and tested with all the cases. You have to prepare those masks before using your models.\n\nAnother point that makes this competition hard is that some segmentation models work with masks that have the same file extension that the images that are used as imput, so we have to invest time transforming from .npy to .png", "votes": 3, "replies": [ { "id": 3310193, "postDate": "2025-11-02T10:48:41.947Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3310194, "postDate": "2025-11-02T10:50:07.043Z", "content": "That's fantastic! You're spot on! Especially for images where objects are clearly segmented into distinct blocks, this approach can indeed be quite effective in distinguishing whether there's a copy-move forgery, given that currently the proportion of authentic images is relatively high.
", "rawMarkdown": "That's fantastic! You're spot on! Especially for images where objects are clearly segmented into distinct blocks, this approach can indeed be quite effective in distinguishing whether there's a copy-move forgery, given that currently the proportion of authentic images is relatively high.", "replies": [ { "id": 3312998, "postDate": "2025-11-08T12:32:04.723Z", "content": "Sorry,I don't catch what you said.Could you explain it a bit?
", "rawMarkdown": "Sorry,I don't catch what you said.Could you explain it a bit?" }, { "id": 3313002, "postDate": "2025-11-08T12:39:56.893Z", "content": "The original poster (OP) means that they hope the organizer can provide masks for \"authentic\" images, rather than just masks for copymove-manipulated images, so as to better train the model. I agree with his/her idea. Specifically, masks could be provided for images that can clearly distinguish objects, and that's it.
", "rawMarkdown": "The original poster (OP) means that they hope the organizer can provide masks for \"authentic\" images, rather than just masks for copymove-manipulated images, so as to better train the model. I agree with his/her idea. Specifically, masks could be provided for images that can clearly distinguish objects, and that's it.\n\n" }, { "id": 3313004, "postDate": "2025-11-08T12:46:57.730Z", "content": "🙏Thank you.
", "rawMarkdown": "🙏Thank you." } ] } ] }, { "id": 3308163, "postDate": "2025-10-28T18:51:25.970Z", "content": "Simple and weak models cannot be used here. We need a good approach and analysis here. I have a bigger problem with GPU availability, I would like to try a couple of models, but I won't have time to train them =(
\n
That makes sense. I also feel like brute-forcing big models isn’t the best path here.\nI think think smarter, more analytical approaches could make a bigger difference than just training large architectures.
", "rawMarkdown": "That makes sense. I also feel like brute-forcing big models isn’t the best path here.\nI think think smarter, more analytical approaches could make a bigger difference than just training large architectures. \n", "votes": 4 } ] }, { "id": 3308161, "postDate": "2025-10-28T18:48:20.033Z", "content": "Maybe it’s the data, maybe the metric.\nWhat do you think makes this challenge difficult. share your insights!
", "rawMarkdown": "Maybe it’s the data, maybe the metric.\nWhat do you think makes this challenge difficult. share your insights!", "votes": 1 } ], "comments": [ { "id": 3310187, "author_name": "Sara R.C.", "author_url": "", "post_date": "2025-11-02T10:12:00.240000", "content": "One thing that I have found is that the authentic images should have empty masks, masks filled with 0 with the same size as the image. Without those masks you can't train a model because you need your model trained and tested with all the cases. You have to prepare those masks before using your models.
\nAnother point that makes this competition hard is that some segmentation models work with masks that have the same file extension that the images that are used as imput, so we have to invest time transforming from .npy to .png
", "votes": 3, "replies": [ { "id": 3310193, "author_name": "", "author_url": "", "post_date": "2025-11-02T10:48:41.947000", "content": "", "votes": 0, "replies": [] }, { "id": 3310194, "author_name": "耶✌", "author_url": "", "post_date": "2025-11-02T10:50:07.043000", "content": "That's fantastic! You're spot on! Especially for images where objects are clearly segmented into distinct blocks, this approach can indeed be quite effective in distinguishing whether there's a copy-move forgery, given that currently the proportion of authentic images is relatively high.
", "votes": 0, "replies": [ { "id": 3312998, "author_name": "howardatri", "author_url": "", "post_date": "2025-11-08T12:32:04.723000", "content": "Sorry,I don't catch what you said.Could you explain it a bit?
", "votes": 0, "replies": [] }, { "id": 3313002, "author_name": "耶✌", "author_url": "", "post_date": "2025-11-08T12:39:56.893000", "content": "The original poster (OP) means that they hope the organizer can provide masks for \"authentic\" images, rather than just masks for copymove-manipulated images, so as to better train the model. I agree with his/her idea. Specifically, masks could be provided for images that can clearly distinguish objects, and that's it.
", "votes": 0, "replies": [] }, { "id": 3313004, "author_name": "howardatri", "author_url": "", "post_date": "2025-11-08T12:46:57.730000", "content": "🙏Thank you.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3308163, "author_name": "Antonoof", "author_url": "", "post_date": "2025-10-28T18:51:25.970000", "content": "Simple and weak models cannot be used here. We need a good approach and analysis here. I have a bigger problem with GPU availability, I would like to try a couple of models, but I won't have time to train them =(
\n
That makes sense. I also feel like brute-forcing big models isn’t the best path here.\nI think think smarter, more analytical approaches could make a bigger difference than just training large architectures.
", "votes": 4, "replies": [] } ] } ], "raw_markdown_by_id": { "3310187": "One thing that I have found is that the authentic images should have empty masks, masks filled with 0 with the same size as the image. Without those masks you can't train a model because you need your model trained and tested with all the cases. You have to prepare those masks before using your models.\n\nAnother point that makes this competition hard is that some segmentation models work with masks that have the same file extension that the images that are used as imput, so we have to invest time transforming from .npy to .png", "3308163": "Simple and weak models cannot be used here. We need a good approach and analysis here. I have a bigger problem with GPU availability, I would like to try a couple of models, but I won't have time to train them =(\n\n", "3308161": "Maybe it’s the data, maybe the metric.\nWhat do you think makes this challenge difficult. share your insights!" } }