Do we need to annotate BOTH the source region (original part that was copied from) AND the target region (copy-pasted part) in our predicted masks to maximize the oF1 score? Or does the evaluation metric primarily reward detection of just the forged/target regions?
\nMy concern is that if the official evaluation only considers the copied/pasted regions (target) for scoring, then spending model capacity on detecting source regions might not be optimal. However, if both regions are required for a high score, then we should focus on improving detection of both areas.
\nHas anyone analyzed the official evaluation code or conducted experiments to determine which approach yields better scores? Any insights would be greatly appreciated!
", "messages": [ { "id": "3313021", "postDate": "11/08/2025 13:44:39", "content": "Do we need to annotate BOTH the source region (original part that was copied from) AND the target region (copy-pasted part) in our predicted masks to maximize the oF1 score? Or does the evaluation metric primarily reward detection of just the forged/target regions?
\nMy concern is that if the official evaluation only considers the copied/pasted regions (target) for scoring, then spending model capacity on detecting source regions might not be optimal. However, if both regions are required for a high score, then we should focus on improving detection of both areas.
\nHas anyone analyzed the official evaluation code or conducted experiments to determine which approach yields better scores? Any insights would be greatly appreciated!
", "rawMarkdown": "Do we need to annotate BOTH the source region (original part that was copied from) AND the target region (copy-pasted part) in our predicted masks to maximize the oF1 score? Or does the evaluation metric primarily reward detection of just the forged/target regions?\n\n**My concern is that if the official evaluation only considers the copied/pasted regions (target) for scoring, then spending model capacity on detecting source regions might not be optimal. However, if both regions are required for a high score, then we should focus on improving detection of both areas.**\n\nHas anyone analyzed the official evaluation code or conducted experiments to determine which approach yields better scores? Any insights would be greatly appreciated!", "votes": null }, { "id": "3313067", "postDate": "11/08/2025 15:37:08", "content": "We need submit both mask,I think
", "rawMarkdown": "We need submit both mask,I think", "votes": null }, { "id": "3315017", "postDate": "11/10/2025 13:06:23", "content": "Yes, that's correct. The goal is to identify both the source of the copy and the pasted region(s).
\nFor each forgery instance (i.e., one source object being copied), we indicate all related regions (both the source and its copies) with the same channel of the ground truth.
\nFor example, if cell A and cell B are both copied in the same image:
\nThe source region for cell A and all its copies are indicated in the ground truth with the same channel (e.g., Channel 1).
The source region for cell B and all its copies are identified with a different channel (e.g., Channel 2).
Good!Thank you very much.
", "rawMarkdown": "Good!Thank you very much.", "votes": null } ], "comments": [ { "id": 3313067, "author_name": "qifeihhh666", "author_url": "", "post_date": "11/08/2025 15:37:08", "content": "We need submit both mask,I think
", "votes": null, "replies": [] }, { "id": 3315017, "author_name": "joophillipecardenuto", "author_url": "", "post_date": "11/10/2025 13:06:23", "content": "Yes, that's correct. The goal is to identify both the source of the copy and the pasted region(s).
\nFor each forgery instance (i.e., one source object being copied), we indicate all related regions (both the source and its copies) with the same channel of the ground truth.
\nFor example, if cell A and cell B are both copied in the same image:
\nThe source region for cell A and all its copies are indicated in the ground truth with the same channel (e.g., Channel 1).
The source region for cell B and all its copies are identified with a different channel (e.g., Channel 2).
Good!Thank you very much.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3313021": "Do we need to annotate BOTH the source region (original part that was copied from) AND the target region (copy-pasted part) in our predicted masks to maximize the oF1 score? Or does the evaluation metric primarily reward detection of just the forged/target regions?\n\n**My concern is that if the official evaluation only considers the copied/pasted regions (target) for scoring, then spending model capacity on detecting source regions might not be optimal. However, if both regions are required for a high score, then we should focus on improving detection of both areas.**\n\nHas anyone analyzed the official evaluation code or conducted experiments to determine which approach yields better scores? Any insights would be greatly appreciated!", "3313067": "We need submit both mask,I think", "3315017": "Yes, that's correct. The goal is to identify both the source of the copy and the pasted region(s).\n\nFor each forgery instance (i.e., one source object being copied), we indicate all related regions (both the source and its copies) with the same channel of the ground truth.\n\nFor example, if cell A and cell B are both copied in the same image:\n\n- The source region for cell A and all its copies are indicated in the ground truth with the same channel (e.g., Channel 1).\n\n- The source region for cell B and all its copies are identified with a different channel (e.g., Channel 2).", "3315119": "Good!Thank you very much." }, "source": "meta" }