{
  "id": 673252,
  "title": "Is Mask-RCNN a valid general approach?",
  "url": "/competitions/recodai-luc-scientific-image-forgery-detection/discussion/673252",
  "author_name": "",
  "post_date": "2026-02-13T12:36:10.443318700Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hello,</p>\n<p>I have been busy training a Mask-RCNN (encoder, classifier+box, mask). I start using a COCO-pretrained Mask R-CNN\n    model = maskrcnn_resnet50_fpn(weights=\"DEFAULT\")\nand did the following:</p>\n<ul>\n<li>Overfit 1 image.</li>\n<li>Overfit 5 images.</li>\n<li>Train on the whole dataset.</li>\n</ul>\n<p>obtaining, on the training dataset, an oF1 = 0.42 +- 0.43\nbut on the test dataset only oF1 = 0.05</p>\n<p>The problem seems to be the classifier, which can't distinguish forged and authentic boxes, and is heavily biased towards \"forged\". Thus, all detected regions are being masked, giving very bad Precision.</p>\n<p>I found out that the encoder has to be trained specifically to find \"forgery\" cues/features (pixel artifacts, abrupt lightning/noise changes etc.). If this is not done, there is no way for the classifier to distinguish the regions. This is my new task.</p>\n<p>I do however wonder if Mask RCNN is a good approach and i should spend more time on this or is there more suitable approaches? Is there pre-trained \"forgery\" encoders that i could just plug in? Or am I in the wrong direction generally? Leave your thoughts!</p>",
  "messages": [
    {
      "id": "3405557",
      "postDate": "02/13/2026 12:36:10",
      "content": "<p>Hello,</p>\n<p>I have been busy training a Mask-RCNN (encoder, classifier+box, mask). I start using a COCO-pretrained Mask R-CNN\n    model = maskrcnn_resnet50_fpn(weights=\"DEFAULT\")\nand did the following:</p>\n<ul>\n<li>Overfit 1 image.</li>\n<li>Overfit 5 images.</li>\n<li>Train on the whole dataset.</li>\n</ul>\n<p>obtaining, on the training dataset, an oF1 = 0.42 +- 0.43\nbut on the test dataset only oF1 = 0.05</p>\n<p>The problem seems to be the classifier, which can't distinguish forged and authentic boxes, and is heavily biased towards \"forged\". Thus, all detected regions are being masked, giving very bad Precision.</p>\n<p>I found out that the encoder has to be trained specifically to find \"forgery\" cues/features (pixel artifacts, abrupt lightning/noise changes etc.). If this is not done, there is no way for the classifier to distinguish the regions. This is my new task.</p>\n<p>I do however wonder if Mask RCNN is a good approach and i should spend more time on this or is there more suitable approaches? Is there pre-trained \"forgery\" encoders that i could just plug in? Or am I in the wrong direction generally? Leave your thoughts!</p>",
      "rawMarkdown": "Hello,\n\nI have been busy training a Mask-RCNN (encoder, classifier+box, mask). I start using a COCO-pretrained Mask R-CNN\n    model = maskrcnn_resnet50_fpn(weights=\"DEFAULT\")\nand did the following:\n- Overfit 1 image.\n- Overfit 5 images.\n- Train on the whole dataset.\n\nobtaining, on the training dataset, an oF1 = 0.42 +- 0.43\nbut on the test dataset only oF1 = 0.05\n\nThe problem seems to be the classifier, which can't distinguish forged and authentic boxes, and is heavily biased towards \"forged\". Thus, all detected regions are being masked, giving very bad Precision.\n\nI found out that the encoder has to be trained specifically to find \"forgery\" cues/features (pixel artifacts, abrupt lightning/noise changes etc.). If this is not done, there is no way for the classifier to distinguish the regions. This is my new task.\n\nI do however wonder if Mask RCNN is a good approach and i should spend more time on this or is there more suitable approaches? Is there pre-trained \"forgery\" encoders that i could just plug in? Or am I in the wrong direction generally? Leave your thoughts!",
      "votes": null
    },
    {
      "id": "3405854",
      "postDate": "02/14/2026 02:56:52",
      "content": "<p>I happened to be taking an online course about DETR when I participated in this competition. According to the course, I think that the DINO series algorithms may outperform the RCNN series on this 10G dataset? I myself am making improvements based on DINOv2, and many open-source solutions seem to have adopted this approach as well.</p>",
      "rawMarkdown": "I happened to be taking an online course about DETR when I participated in this competition. According to the course, I think that the DINO series algorithms may outperform the RCNN series on this 10G dataset? I myself am making improvements based on DINOv2, and many open-source solutions seem to have adopted this approach as well.",
      "votes": null
    },
    {
      "id": "3405856",
      "postDate": "02/14/2026 02:59:34",
      "content": "<p>May God bless all overfitters🥹</p>",
      "rawMarkdown": "May God bless all overfitters🥹",
      "votes": null
    },
    {
      "id": "3406116",
      "postDate": "02/14/2026 16:49:26",
      "content": "<p>Hello and thanks!! I think I might try DINOv2 as well, but i find so strange that the \"EDA and RCNN\" notebook got so many votes. After all it only has a score of .303, the same as \"all authentic\" prediction 🤷‍♂️</p>",
      "rawMarkdown": "Hello and thanks!! I think I might try DINOv2 as well, but i find so strange that the \"EDA and RCNN\" notebook got so many votes. After all it only has a score of .303, the same as \"all authentic\" prediction 🤷‍♂️",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3405854,
      "author_name": "opencv411",
      "author_url": "",
      "post_date": "02/14/2026 02:56:52",
      "content": "<p>I happened to be taking an online course about DETR when I participated in this competition. According to the course, I think that the DINO series algorithms may outperform the RCNN series on this 10G dataset? I myself am making improvements based on DINOv2, and many open-source solutions seem to have adopted this approach as well.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3405856,
          "author_name": "opencv411",
          "author_url": "",
          "post_date": "02/14/2026 02:59:34",
          "content": "<p>May God bless all overfitters🥹</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 3406116,
          "author_name": "andreshzapke",
          "author_url": "",
          "post_date": "02/14/2026 16:49:26",
          "content": "<p>Hello and thanks!! I think I might try DINOv2 as well, but i find so strange that the \"EDA and RCNN\" notebook got so many votes. After all it only has a score of .303, the same as \"all authentic\" prediction 🤷‍♂️</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3405557": "Hello,\n\nI have been busy training a Mask-RCNN (encoder, classifier+box, mask). I start using a COCO-pretrained Mask R-CNN\n    model = maskrcnn_resnet50_fpn(weights=\"DEFAULT\")\nand did the following:\n- Overfit 1 image.\n- Overfit 5 images.\n- Train on the whole dataset.\n\nobtaining, on the training dataset, an oF1 = 0.42 +- 0.43\nbut on the test dataset only oF1 = 0.05\n\nThe problem seems to be the classifier, which can't distinguish forged and authentic boxes, and is heavily biased towards \"forged\". Thus, all detected regions are being masked, giving very bad Precision.\n\nI found out that the encoder has to be trained specifically to find \"forgery\" cues/features (pixel artifacts, abrupt lightning/noise changes etc.). If this is not done, there is no way for the classifier to distinguish the regions. This is my new task.\n\nI do however wonder if Mask RCNN is a good approach and i should spend more time on this or is there more suitable approaches? Is there pre-trained \"forgery\" encoders that i could just plug in? Or am I in the wrong direction generally? Leave your thoughts!",
    "3405854": "I happened to be taking an online course about DETR when I participated in this competition. According to the course, I think that the DINO series algorithms may outperform the RCNN series on this 10G dataset? I myself am making improvements based on DINOv2, and many open-source solutions seem to have adopted this approach as well.",
    "3405856": "May God bless all overfitters🥹",
    "3406116": "Hello and thanks!! I think I might try DINOv2 as well, but i find so strange that the \"EDA and RCNN\" notebook got so many votes. After all it only has a score of .303, the same as \"all authentic\" prediction 🤷‍♂️"
  },
  "source": "meta"
}