If test set would have pairs, then it might introduce some kind of a leakage. We can use that information to alter the predictions. For example, we can store all image embeddings and their predictions. When a new image that is very similar to a previous one comes, we can check that previous image's prediction and predict the current one based on whether the previous one was authentic or forged.
", "messages": [ { "id": "3378150", "postDate": "12/17/2025 16:02:11", "content": "If test set would have pairs, then it might introduce some kind of a leakage. We can use that information to alter the predictions. For example, we can store all image embeddings and their predictions. When a new image that is very similar to a previous one comes, we can check that previous image's prediction and predict the current one based on whether the previous one was authentic or forged.
", "rawMarkdown": "If test set would have pairs, then it might introduce some kind of a leakage. We can use that information to alter the predictions. For example, we can store all image embeddings and their predictions. When a new image that is very similar to a previous one comes, we can check that previous image's prediction and predict the current one based on whether the previous one was authentic or forged.", "votes": null }, { "id": "3378160", "postDate": "12/17/2025 16:42:12", "content": "I don't think so given the information provided.
\nhttps://www.kaggle.com/competitions/recodai-luc-scientific-image-forgery-detection/discussion/617968
\nFrom the discussion there:
\n\n\nThese new images were acquired and labeled using the exact process that will be used for the final test.
\n
And:
\n\n\nFor the supplemental and test data, we collected them from actual reported cases of scientific image manipulation. They reflect real-world complexity and may include multi-panel figures and single-panel images.
\n
Given that the authentic in test set will likely be real non forgery images from publications and the forgeries will be real discovered forgeries (ie they don't have the authentic for them to create a pair).
", "rawMarkdown": "I don't think so given the information provided.\n\nhttps://www.kaggle.com/competitions/recodai-luc-scientific-image-forgery-detection/discussion/617968\n\nFrom the discussion there:\n\n> These new images were acquired and labeled using the exact process that will be used for the final test.\n\nAnd:\n\n> For the supplemental and test data, we collected them from actual reported cases of scientific image manipulation. They reflect real-world complexity and may include multi-panel figures and single-panel images.\n\nGiven that the authentic in test set will likely be real non forgery images from publications and the forgeries will be real discovered forgeries (ie they don't have the authentic for them to create a pair).", "votes": null } ], "comments": [ { "id": 3378160, "author_name": "timothylovett", "author_url": "", "post_date": "12/17/2025 16:42:12", "content": "I don't think so given the information provided.
\nhttps://www.kaggle.com/competitions/recodai-luc-scientific-image-forgery-detection/discussion/617968
\nFrom the discussion there:
\n\n\nThese new images were acquired and labeled using the exact process that will be used for the final test.
\n
And:
\n\n\nFor the supplemental and test data, we collected them from actual reported cases of scientific image manipulation. They reflect real-world complexity and may include multi-panel figures and single-panel images.
\n
Given that the authentic in test set will likely be real non forgery images from publications and the forgeries will be real discovered forgeries (ie they don't have the authentic for them to create a pair).
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3378150": "If test set would have pairs, then it might introduce some kind of a leakage. We can use that information to alter the predictions. For example, we can store all image embeddings and their predictions. When a new image that is very similar to a previous one comes, we can check that previous image's prediction and predict the current one based on whether the previous one was authentic or forged.", "3378160": "I don't think so given the information provided.\n\nhttps://www.kaggle.com/competitions/recodai-luc-scientific-image-forgery-detection/discussion/617968\n\nFrom the discussion there:\n\n> These new images were acquired and labeled using the exact process that will be used for the final test.\n\nAnd:\n\n> For the supplemental and test data, we collected them from actual reported cases of scientific image manipulation. They reflect real-world complexity and may include multi-panel figures and single-panel images.\n\nGiven that the authentic in test set will likely be real non forgery images from publications and the forgeries will be real discovered forgeries (ie they don't have the authentic for them to create a pair)." }, "source": "meta" }