I would like to thank Kaggle and Recod.ai/LUC for hosting this interesting competition. I finished 5th on the private leaderboard.
\nFinal scores:
\nMy solution focused on simplicity and robustness. I treated the task primarily as a semantic segmentation problem.
\nFor training mask generation, when an image contained multiple copy-move instances, I used the union of all manipulated regions as the training target. In other words, the model was trained as a binary manipulated-region segmenter rather than as an instance segmentation model.
\nAt inference time, I also did not separate multiple copy-move instances. This was a deliberate simplification.
\nTwo factors contributed most significantly to the final score:
\nI mainly validated my models on the 48 supplemental images provided by the host, because these multi-panel images were the closest match to the actual test set.
\nI used pretrained DINOv2/DINOv3 backbones with a very simple segmentation head.
\nThe model takes the final patch tokens from the ViT backbone, reshapes them into a 2D feature grid, and applies a single 1×1 Conv2d head to produce one logit map.
\nDuring training, I downsampled the ground-truth mask to match the patch-grid resolution:
\nDuring inference, after predicting the low-resolution mask, I simply resized it back to the original image size using bilinear interpolation.
\nThe most useful augmentations were simple image-level augmentations:
\nA.RandomRotate90(p=1.0)\nA.HorizontalFlip(p=0.5)\nA.GaussianBlur(p=0.15)\nA.ToGray(...)\nA.HueSaturationValue(...)\nA.RandomBrightnessContrast(p=0.9)\nA.ImageCompression(p=0.25)\nI used fairly aggressive color augmentation because the manipulation signal is not always color-specific. The model needed to learn structural and local consistency cues rather than memorize color distributions.
\nrotate90 test-time augmentationThe post-processing pipeline was simple:
\nmin_area, classify the sample as authentic and stop.mask_thr.I started with a DINOv2 model. Its validation score on the supplemental images was only 0.092, although the validation score on a split of the provided training set was much higher.
\nAfter inspecting the training data, I found that although the dataset was not small, it contained only a limited variety of image types. As a result, the fine-tuned model did not generalize well to the supplemental set, which was visually quite different from the training set.
\nTo improve generalization, I searched for external datasets that could increase the diversity of the training data. I eventually added the following datasets:
\n\nValidation results on the supplemental images were:
\n| Training data | \nValidation score | \n
|---|---|
| Competition training set only | \n0.092 | \n
| Competition training set + CASIA | \n0.185 | \n
| Competition training set + GRIP | \n0.095 | \n
| Competition training set + FAU | \n0.129 | \n
| Competition training set + CASIA + GRIP + FAU | \n0.240 | \n
Although these external datasets are not biomedical, their diversity helped the model learn more general copy-move detection cues.
\nBecause copy-move detection is a difficult task, stronger pretrained backbones made a large difference.
\nUpgrading the backbone from DINOv2 Giant to DINOv3 Huge+ improved the validation score from the earlier range to around 0.35. Since these two models are relatively close in size, I believe the improvement mainly came from the stronger pretraining of DINOv3.
\nI then used the original DINOv3 7B model, which further improved the validation score to 0.51. With additional hyperparameter tuning, the best validation score exceeded 0.56.
\nOverall, the final solution remained simple: a strong pretrained ViT backbone, a minimal segmentation head, diverse training data, and lightweight post-processing.
", "messages": [ { "id": 3454335, "postDate": "2026-05-06T23:20:47.997Z", "content": "I would like to thank Kaggle and Recod.ai/LUC for hosting this interesting competition. I finished 5th on the private leaderboard.
\nFinal scores:
\nMy solution focused on simplicity and robustness. I treated the task primarily as a semantic segmentation problem.
\nFor training mask generation, when an image contained multiple copy-move instances, I used the union of all manipulated regions as the training target. In other words, the model was trained as a binary manipulated-region segmenter rather than as an instance segmentation model.
\nAt inference time, I also did not separate multiple copy-move instances. This was a deliberate simplification.
\nTwo factors contributed most significantly to the final score:
\nI mainly validated my models on the 48 supplemental images provided by the host, because these multi-panel images were the closest match to the actual test set.
\nI used pretrained DINOv2/DINOv3 backbones with a very simple segmentation head.
\nThe model takes the final patch tokens from the ViT backbone, reshapes them into a 2D feature grid, and applies a single 1×1 Conv2d head to produce one logit map.
\nDuring training, I downsampled the ground-truth mask to match the patch-grid resolution:
\nDuring inference, after predicting the low-resolution mask, I simply resized it back to the original image size using bilinear interpolation.
\nThe most useful augmentations were simple image-level augmentations:
\nA.RandomRotate90(p=1.0)\nA.HorizontalFlip(p=0.5)\nA.GaussianBlur(p=0.15)\nA.ToGray(...)\nA.HueSaturationValue(...)\nA.RandomBrightnessContrast(p=0.9)\nA.ImageCompression(p=0.25)\nI used fairly aggressive color augmentation because the manipulation signal is not always color-specific. The model needed to learn structural and local consistency cues rather than memorize color distributions.
\nrotate90 test-time augmentationThe post-processing pipeline was simple:
\nmin_area, classify the sample as authentic and stop.mask_thr.I started with a DINOv2 model. Its validation score on the supplemental images was only 0.092, although the validation score on a split of the provided training set was much higher.
\nAfter inspecting the training data, I found that although the dataset was not small, it contained only a limited variety of image types. As a result, the fine-tuned model did not generalize well to the supplemental set, which was visually quite different from the training set.
\nTo improve generalization, I searched for external datasets that could increase the diversity of the training data. I eventually added the following datasets:
\n\nValidation results on the supplemental images were:
\n| Training data | \nValidation score | \n
|---|---|
| Competition training set only | \n0.092 | \n
| Competition training set + CASIA | \n0.185 | \n
| Competition training set + GRIP | \n0.095 | \n
| Competition training set + FAU | \n0.129 | \n
| Competition training set + CASIA + GRIP + FAU | \n0.240 | \n
Although these external datasets are not biomedical, their diversity helped the model learn more general copy-move detection cues.
\nBecause copy-move detection is a difficult task, stronger pretrained backbones made a large difference.
\nUpgrading the backbone from DINOv2 Giant to DINOv3 Huge+ improved the validation score from the earlier range to around 0.35. Since these two models are relatively close in size, I believe the improvement mainly came from the stronger pretraining of DINOv3.
\nI then used the original DINOv3 7B model, which further improved the validation score to 0.51. With additional hyperparameter tuning, the best validation score exceeded 0.56.
\nOverall, the final solution remained simple: a strong pretrained ViT backbone, a minimal segmentation head, diverse training data, and lightweight post-processing.
", "rawMarkdown": "I would like to thank Kaggle and Recod.ai/LUC for hosting this interesting competition. I finished 5th on the private leaderboard.\n\n**Final scores:**\n\n* Public LB: **0.422**\n* Private LB: **0.376**\n* [**inference code**](https://www.kaggle.com/code/wowfattie/recodfinal)\n* **training code attached at the end**\n\n## Overview\n\nMy solution focused on simplicity and robustness. I treated the task primarily as a semantic segmentation problem.\n\nFor training mask generation, when an image contained multiple copy-move instances, I used the union of all manipulated regions as the training target. In other words, the model was trained as a binary manipulated-region segmenter rather than as an instance segmentation model.\n\nAt inference time, I also did not separate multiple copy-move instances. This was a deliberate simplification.\n\nTwo factors contributed most significantly to the final score:\n\n1. Using external datasets to improve generalization.\n2. Using the strongest available pretrained DINOv3 backbone.\n\n## Validation\n\nI mainly validated my models on the 48 supplemental images provided by the host, because these multi-panel images were the closest match to the actual test set.\n\n## Backbone and Segmentation Head\n\nI used pretrained DINOv2/DINOv3 backbones with a very simple segmentation head.\n\nThe model takes the final patch tokens from the ViT backbone, reshapes them into a 2D feature grid, and applies a single 1×1 Conv2d head to produce one logit map.\n\nDuring training, I downsampled the ground-truth mask to match the patch-grid resolution:\n\n* 14×14 for DINOv2\n* 16×16 for DINOv3\n\nDuring inference, after predicting the low-resolution mask, I simply resized it back to the original image size using bilinear interpolation.\n\n## Augmentations\n\nThe most useful augmentations were simple image-level augmentations:\n\n```python\nA.RandomRotate90(p=1.0)\nA.HorizontalFlip(p=0.5)\nA.GaussianBlur(p=0.15)\nA.ToGray(...)\nA.HueSaturationValue(...)\nA.RandomBrightnessContrast(p=0.9)\nA.ImageCompression(p=0.25)\n```\n\nI used fairly aggressive color augmentation because the manipulation signal is not always color-specific. The model needed to learn structural and local consistency cues rather than memorize color distributions.\n\n## Important Training Choices\n\n* BF16 autocast\n* Full fine-tuning\n* FlashAttention-2\n* Input image size: 1024 × 1024\n* Vanilla BCE loss\n\n## Important Inference Choices\n\n* 8-bit inference to fit within the 16GB memory limit of a T4 GPU\n* SDPA attention\n* Input image size: 1024 × 1024\n* Four-way `rotate90` test-time augmentation\n\n## Post-processing\n\nThe post-processing pipeline was simple:\n\n1. Create a high-confidence binary mask using a fixed threshold.\n2. Compute the area of this high-confidence mask by summing its active pixels.\n3. If the high-confidence area is smaller than `min_area`, classify the sample as authentic and stop.\n4. Create the final binary mask using the configured threshold `mask_thr`.\n5. If the final mask contains no active pixels, classify the sample as authentic and stop.\n\n## Early Results\n\nI started with a DINOv2 model. Its validation score on the supplemental images was only **0.092**, although the validation score on a split of the provided training set was much higher.\n\nAfter inspecting the training data, I found that although the dataset was not small, it contained only a limited variety of image types. As a result, the fine-tuned model did not generalize well to the supplemental set, which was visually quite different from the training set.\n\n## External Datasets\n\nTo improve generalization, I searched for external datasets that could increase the diversity of the training data. I eventually added the following datasets:\n\n* [CASIA](https://www.kaggle.com/datasets/divg07/casia-20-image-tampering-detection-dataset)\n* [GRIP](https://www.grip.unina.it/download/prog/CMFD/)\n* [FAU](https://www.cs1.tf.fau.de/research/multimedia-security/code/image-manipulation-dataset/#collapse_1)\n\nValidation results on the supplemental images were:\n\n| Training data | Validation score |\n| --------------------------------------------- | ---------------: |\n| Competition training set only | 0.092 |\n| Competition training set + CASIA | 0.185 |\n| Competition training set + GRIP | 0.095 |\n| Competition training set + FAU | 0.129 |\n| Competition training set + CASIA + GRIP + FAU | 0.240 |\n\nAlthough these external datasets are not biomedical, their diversity helped the model learn more general copy-move detection cues.\n\n## Better Backbones\n\nBecause copy-move detection is a difficult task, stronger pretrained backbones made a large difference.\n\nUpgrading the backbone from **DINOv2 Giant** to **DINOv3 Huge+** improved the validation score from the earlier range to around **0.35**. Since these two models are relatively close in size, I believe the improvement mainly came from the stronger pretraining of DINOv3.\n\nI then used the original **DINOv3 7B** model, which further improved the validation score to **0.51**. With additional hyperparameter tuning, the best validation score exceeded **0.56**.\n\nOverall, the final solution remained simple: a strong pretrained ViT backbone, a minimal segmentation head, diverse training data, and lightweight post-processing.\n", "votes": 5 } ], "comments": [], "raw_markdown_by_id": { "3454335": "I would like to thank Kaggle and Recod.ai/LUC for hosting this interesting competition. I finished 5th on the private leaderboard.\n\n**Final scores:**\n\n* Public LB: **0.422**\n* Private LB: **0.376**\n* [**inference code**](https://www.kaggle.com/code/wowfattie/recodfinal)\n* **training code attached at the end**\n\n## Overview\n\nMy solution focused on simplicity and robustness. I treated the task primarily as a semantic segmentation problem.\n\nFor training mask generation, when an image contained multiple copy-move instances, I used the union of all manipulated regions as the training target. In other words, the model was trained as a binary manipulated-region segmenter rather than as an instance segmentation model.\n\nAt inference time, I also did not separate multiple copy-move instances. This was a deliberate simplification.\n\nTwo factors contributed most significantly to the final score:\n\n1. Using external datasets to improve generalization.\n2. Using the strongest available pretrained DINOv3 backbone.\n\n## Validation\n\nI mainly validated my models on the 48 supplemental images provided by the host, because these multi-panel images were the closest match to the actual test set.\n\n## Backbone and Segmentation Head\n\nI used pretrained DINOv2/DINOv3 backbones with a very simple segmentation head.\n\nThe model takes the final patch tokens from the ViT backbone, reshapes them into a 2D feature grid, and applies a single 1×1 Conv2d head to produce one logit map.\n\nDuring training, I downsampled the ground-truth mask to match the patch-grid resolution:\n\n* 14×14 for DINOv2\n* 16×16 for DINOv3\n\nDuring inference, after predicting the low-resolution mask, I simply resized it back to the original image size using bilinear interpolation.\n\n## Augmentations\n\nThe most useful augmentations were simple image-level augmentations:\n\n```python\nA.RandomRotate90(p=1.0)\nA.HorizontalFlip(p=0.5)\nA.GaussianBlur(p=0.15)\nA.ToGray(...)\nA.HueSaturationValue(...)\nA.RandomBrightnessContrast(p=0.9)\nA.ImageCompression(p=0.25)\n```\n\nI used fairly aggressive color augmentation because the manipulation signal is not always color-specific. The model needed to learn structural and local consistency cues rather than memorize color distributions.\n\n## Important Training Choices\n\n* BF16 autocast\n* Full fine-tuning\n* FlashAttention-2\n* Input image size: 1024 × 1024\n* Vanilla BCE loss\n\n## Important Inference Choices\n\n* 8-bit inference to fit within the 16GB memory limit of a T4 GPU\n* SDPA attention\n* Input image size: 1024 × 1024\n* Four-way `rotate90` test-time augmentation\n\n## Post-processing\n\nThe post-processing pipeline was simple:\n\n1. Create a high-confidence binary mask using a fixed threshold.\n2. Compute the area of this high-confidence mask by summing its active pixels.\n3. If the high-confidence area is smaller than `min_area`, classify the sample as authentic and stop.\n4. Create the final binary mask using the configured threshold `mask_thr`.\n5. If the final mask contains no active pixels, classify the sample as authentic and stop.\n\n## Early Results\n\nI started with a DINOv2 model. Its validation score on the supplemental images was only **0.092**, although the validation score on a split of the provided training set was much higher.\n\nAfter inspecting the training data, I found that although the dataset was not small, it contained only a limited variety of image types. As a result, the fine-tuned model did not generalize well to the supplemental set, which was visually quite different from the training set.\n\n## External Datasets\n\nTo improve generalization, I searched for external datasets that could increase the diversity of the training data. I eventually added the following datasets:\n\n* [CASIA](https://www.kaggle.com/datasets/divg07/casia-20-image-tampering-detection-dataset)\n* [GRIP](https://www.grip.unina.it/download/prog/CMFD/)\n* [FAU](https://www.cs1.tf.fau.de/research/multimedia-security/code/image-manipulation-dataset/#collapse_1)\n\nValidation results on the supplemental images were:\n\n| Training data | Validation score |\n| --------------------------------------------- | ---------------: |\n| Competition training set only | 0.092 |\n| Competition training set + CASIA | 0.185 |\n| Competition training set + GRIP | 0.095 |\n| Competition training set + FAU | 0.129 |\n| Competition training set + CASIA + GRIP + FAU | 0.240 |\n\nAlthough these external datasets are not biomedical, their diversity helped the model learn more general copy-move detection cues.\n\n## Better Backbones\n\nBecause copy-move detection is a difficult task, stronger pretrained backbones made a large difference.\n\nUpgrading the backbone from **DINOv2 Giant** to **DINOv3 Huge+** improved the validation score from the earlier range to around **0.35**. Since these two models are relatively close in size, I believe the improvement mainly came from the stronger pretraining of DINOv3.\n\nI then used the original **DINOv3 7B** model, which further improved the validation score to **0.51**. With additional hyperparameter tuning, the best validation score exceeded **0.56**.\n\nOverall, the final solution remained simple: a strong pretrained ViT backbone, a minimal segmentation head, diverse training data, and lightweight post-processing.\n" } }