{ "id": 641008, "title": "Using Segment Anything (SAM2) Masks to Generate Forgery Candidates", "url": "/competitions/recodai-luc-scientific-image-forgery-detection/discussion/641008", "author_name": "", "post_date": "2025-11-26T12:55:29.753619300Z", "votes": 3, "comment_count": 7, "views": 0, "content": "

The key challenge in digital forgery detection is efficiently identifying potential candidate regions. I'm proposing a novel approach: using the Segment Anything Model (SAM) to generate high-confidence, object-level forgery candidates based on the idea that most attacks involve manipulating distinct objects.

\n

This strategy uses SAM's segmentation power to drastically narrow the search space for a final detection model.

\n

Resources for Discussion & Collaboration

\n

To start the discussion and modeling, I've shared the following:

\n
    \n

  1. EDA + Visualization Kernel:

    \n
    2. \n

  2. Generated Dataset (Images + Masks):

    \n
    3. \n

  3. SAM Generation Kernel:

    \n
    4. \n
\n

Let's discuss and optimize this approach for powerful detection!

", "messages": [ { "id": "3349109", "postDate": "11/26/2025 12:55:29", "content": "

The key challenge in digital forgery detection is efficiently identifying potential candidate regions. I'm proposing a novel approach: using the Segment Anything Model (SAM) to generate high-confidence, object-level forgery candidates based on the idea that most attacks involve manipulating distinct objects.

\n

This strategy uses SAM's segmentation power to drastically narrow the search space for a final detection model.

\n

Resources for Discussion & Collaboration

\n

To start the discussion and modeling, I've shared the following:

\n
    \n

  1. EDA + Visualization Kernel:

    \n
    2. \n

  2. Generated Dataset (Images + Masks):

    \n
    3. \n

  3. SAM Generation Kernel:

    \n
    4. \n
\n

Let's discuss and optimize this approach for powerful detection!

", "rawMarkdown": "The key challenge in digital forgery detection is efficiently identifying potential candidate regions. I'm proposing a novel approach: using the **Segment Anything Model (SAM) to generate high-confidence, object-level forgery candidates** based on the idea that most attacks involve manipulating distinct objects.\n\nThis strategy uses SAM's segmentation power to drastically narrow the search space for a final detection model.\n\n## Resources for Discussion & Collaboration\n\nTo start the discussion and modeling, I've shared the following:\n\n1. **EDA + Visualization Kernel:**\n * **Link:** \n * *What it shows:* Visualization of SAM-generated masks and their overlap with ground truth forged regions.\n\n2. **Generated Dataset (Images + Masks):**\n * **Link:** \n * *What it is:* The full dataset of images and corresponding SAM2 segmented masks (`.npz` format).\n\n3. **SAM Generation Kernel:**\n * **Link:** \n * *What it does:* The kernel used to generate the mask dataset.\n\nLet's discuss and optimize this approach for powerful detection!", "votes": null }, { "id": "3351678", "postDate": "11/28/2025 15:12:49", "content": "

I took this approach a couple weeks ago but was not able to beat the default all “authentic” submission. It may be worth revisiting with Segment Anything 3 which just released

", "rawMarkdown": "I took this approach a couple weeks ago but was not able to beat the default all “authentic” submission. It may be worth revisiting with Segment Anything 3 which just released", "votes": null }, { "id": "3351731", "postDate": "11/28/2025 15:43:27", "content": "

so you think that the candidates are weak? Also, how did it perform on the validation set?

", "rawMarkdown": "so you think that the candidates are weak? Also, how did it perform on the validation set?", "votes": null }, { "id": "3352026", "postDate": "11/28/2025 20:17:16", "content": "

validation F1 score of 0.29

", "rawMarkdown": "validation F1 score of 0.29", "votes": null }, { "id": "3380894", "postDate": "12/23/2025 09:41:11", "content": "

The problem is many copy-moves involve replacing objects with the background to hide them, so it won't work. You need something that generates copy-move candidates including Background Patches + a refining step to get pixel-perfect boundaries

", "rawMarkdown": "The problem is many copy-moves involve replacing objects with the background to hide them, so it won't work. You need something that generates copy-move candidates including Background Patches + a refining step to get pixel-perfect boundaries", "votes": null }, { "id": "3380899", "postDate": "12/23/2025 10:08:56", "content": "

That is a good call, I think a good initial version would be just identify copied foreground objects..

", "rawMarkdown": "That is a good call, I think a good initial version would be just identify copied foreground objects..", "votes": null }, { "id": "3380908", "postDate": "12/23/2025 10:41:50", "content": "

Yes, you can try using object detector from the recodai discord, they used it to make this forgery dataset automatically

", "rawMarkdown": "Yes, you can try using object detector from the recodai discord, they used it to make this forgery dataset automatically", "votes": null }, { "id": "3380920", "postDate": "12/23/2025 11:10:04", "content": "

Yes I know…

", "rawMarkdown": "Yes I know...", "votes": null } ], "comments": [ { "id": 3351678, "author_name": "returnofsputnik", "author_url": "", "post_date": "11/28/2025 15:12:49", "content": "

I took this approach a couple weeks ago but was not able to beat the default all “authentic” submission. It may be worth revisiting with Segment Anything 3 which just released

", "votes": null, "replies": [ { "id": 3351731, "author_name": "jirkaborovec", "author_url": "", "post_date": "11/28/2025 15:43:27", "content": "

so you think that the candidates are weak? Also, how did it perform on the validation set?

", "votes": null, "replies": [ { "id": 3352026, "author_name": "returnofsputnik", "author_url": "", "post_date": "11/28/2025 20:17:16", "content": "

validation F1 score of 0.29

", "votes": null, "replies": [] } ] } ] }, { "id": 3380894, "author_name": "theodorlu", "author_url": "", "post_date": "12/23/2025 09:41:11", "content": "

The problem is many copy-moves involve replacing objects with the background to hide them, so it won't work. You need something that generates copy-move candidates including Background Patches + a refining step to get pixel-perfect boundaries

", "votes": null, "replies": [ { "id": 3380899, "author_name": "jirkaborovec", "author_url": "", "post_date": "12/23/2025 10:08:56", "content": "

That is a good call, I think a good initial version would be just identify copied foreground objects..

", "votes": null, "replies": [ { "id": 3380908, "author_name": "theodorlu", "author_url": "", "post_date": "12/23/2025 10:41:50", "content": "

Yes, you can try using object detector from the recodai discord, they used it to make this forgery dataset automatically

", "votes": null, "replies": [ { "id": 3380920, "author_name": "jirkaborovec", "author_url": "", "post_date": "12/23/2025 11:10:04", "content": "

Yes I know…

", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "3349109": "The key challenge in digital forgery detection is efficiently identifying potential candidate regions. I'm proposing a novel approach: using the **Segment Anything Model (SAM) to generate high-confidence, object-level forgery candidates** based on the idea that most attacks involve manipulating distinct objects.\n\nThis strategy uses SAM's segmentation power to drastically narrow the search space for a final detection model.\n\n## Resources for Discussion & Collaboration\n\nTo start the discussion and modeling, I've shared the following:\n\n1. **EDA + Visualization Kernel:**\n * **Link:** \n * *What it shows:* Visualization of SAM-generated masks and their overlap with ground truth forged regions.\n\n2. **Generated Dataset (Images + Masks):**\n * **Link:** \n * *What it is:* The full dataset of images and corresponding SAM2 segmented masks (`.npz` format).\n\n3. **SAM Generation Kernel:**\n * **Link:** \n * *What it does:* The kernel used to generate the mask dataset.\n\nLet's discuss and optimize this approach for powerful detection!", "3351678": "I took this approach a couple weeks ago but was not able to beat the default all “authentic” submission. It may be worth revisiting with Segment Anything 3 which just released", "3351731": "so you think that the candidates are weak? Also, how did it perform on the validation set?", "3352026": "validation F1 score of 0.29", "3380894": "The problem is many copy-moves involve replacing objects with the background to hide them, so it won't work. You need something that generates copy-move candidates including Background Patches + a refining step to get pixel-perfect boundaries", "3380899": "That is a good call, I think a good initial version would be just identify copied foreground objects..", "3380908": "Yes, you can try using object detector from the recodai discord, they used it to make this forgery dataset automatically", "3380920": "Yes I know..." }, "source": "meta" }