![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6102605%2F2e5be21c54cb7bf3b304fea61383b182%2F2026-04-29%20002935.png?generation=1777390204319944&alt=media\")
First of all, thank you to the organizers for hosting such a wonderful competition. It was a highly meaningful theme—protecting the integrity of scientific research—and I learned a great deal from it.
\nResult: Public 3rd → Private 2nd
\nRather than a deep learning-based approach, my solution is built around classical feature matching (SIFT + RANSAC), combined with YOLO-based preprocessing and custom post-processing in a 3-stage pipeline.
\nStage 1: Valid Region Detection (Panel/Text detection with YOLOv8)\n ↓\nStage 2: Feature Matching (SIFT + G2NN + RANSAC)\n ↓\nStage 3: Post-processing (Cluster merging + High-precision mask generation)\nIn the subsequent feature matching stage, graph regions produce a large number of false positives, and text regions cause erroneous matching on caption parts (e.g., matching on the \"mm\" portion of \"10 mm\"). These lead to significant accuracy degradation, making their exclusion essential.
\nI trained a Panel detection model and a Text detection model using YOLOv8-m to address this problem.
\nTo exclude irrelevant regions such as graphs, only relevant areas (image regions) are detected. Simultaneously, YOLO classifies detections into 3 classes: corn images, Western blot protein images, and other biological images. The reason for separating classes is that the optimal RANSAC inlier threshold differs by image type (described below).
\nSince no suitable training data existed, I addressed this by automatically generating synthetic data.
\nExisting OCR solutions (EasyOCR, etc.) were insufficiently accurate—for example, misrecognizing round cells as the letter \"O\". Since text in scientific paper images has minimal distortion, I determined that OCR-specific heads were unnecessary and trained YOLOv8-m as an object detector.
\nBiological images have very little texture, and default settings fail to extract sufficient keypoints.
\ncontrast_threshold=0.001 (approximately 1/40 of the default 0.04) to extract a large number of keypointsThe absolute L2 distance values vary greatly across images, making fixed thresholds impractical.
\nAdopted **G2NN **: sort matching distances and accept only matches up to the breakpoint where $T_{n+1}/T_n > \\alpha$ ($\\alpha=0.7$). This achieves threshold-independent matching.
\nDue to the enormous number of keypoints, used FLANN KDTree approximate nearest neighbor search to reduce complexity from $O(N^2)$ to $O(N \\log N)$.
\nTwo sampling strategies were employed:
\nAdditionally, RANSAC inlier thresholds were adjusted per image type based on the 3 classes detected in Stage 1:
\nAfter feature matching, I discovered that careful clustering is essential. Without it, the F1 score (the competition metric) drops significantly. To achieve high-precision clustering, I implemented the following processing.
\n| Component | \nDetails | \n
|---|---|
| Panel Detection | \nYOLOv8-m × 5 (WBF ensemble), trained on synthetic data | \n
| Text Detection | \nYOLOv8-m, manual annotation + synthetic data | \n
| Features | \nSIFT (contrast_threshold=0.001) | \n
| Matching | \nG2NN (α=0.7) + FLANN KDTree | \n
| Geometric Estimation | \nAffine RANSAC + H matrix filtering | \n
| Clustering | \nHDBSCAN + 2-stage merging (H similarity → IoMin) | \n
| Mask Refinement | \nConvex hull + H transformation error-based refinement | \n
Code is available here https://github.com/oshima-yoppi/luc_pub.
", "messages": [ { "id": "3448118", "postDate": "04/24/2026 17:14:08", "content": "First of all, thank you to the organizers for hosting such a wonderful competition. It was a highly meaningful theme—protecting the integrity of scientific research—and I learned a great deal from it.
\nResult: Public 3rd → Private 2nd
\nRather than a deep learning-based approach, my solution is built around classical feature matching (SIFT + RANSAC), combined with YOLO-based preprocessing and custom post-processing in a 3-stage pipeline.
\nStage 1: Valid Region Detection (Panel/Text detection with YOLOv8)\n ↓\nStage 2: Feature Matching (SIFT + G2NN + RANSAC)\n ↓\nStage 3: Post-processing (Cluster merging + High-precision mask generation)\nIn the subsequent feature matching stage, graph regions produce a large number of false positives, and text regions cause erroneous matching on caption parts (e.g., matching on the \"mm\" portion of \"10 mm\"). These lead to significant accuracy degradation, making their exclusion essential.
\nI trained a Panel detection model and a Text detection model using YOLOv8-m to address this problem.
\nTo exclude irrelevant regions such as graphs, only relevant areas (image regions) are detected. Simultaneously, YOLO classifies detections into 3 classes: corn images, Western blot protein images, and other biological images. The reason for separating classes is that the optimal RANSAC inlier threshold differs by image type (described below).
\nSince no suitable training data existed, I addressed this by automatically generating synthetic data.
\nExisting OCR solutions (EasyOCR, etc.) were insufficiently accurate—for example, misrecognizing round cells as the letter \"O\". Since text in scientific paper images has minimal distortion, I determined that OCR-specific heads were unnecessary and trained YOLOv8-m as an object detector.
\nBiological images have very little texture, and default settings fail to extract sufficient keypoints.
\ncontrast_threshold=0.001 (approximately 1/40 of the default 0.04) to extract a large number of keypointsThe absolute L2 distance values vary greatly across images, making fixed thresholds impractical.
\nAdopted **G2NN **: sort matching distances and accept only matches up to the breakpoint where $T_{n+1}/T_n > \\alpha$ ($\\alpha=0.7$). This achieves threshold-independent matching.
\nDue to the enormous number of keypoints, used FLANN KDTree approximate nearest neighbor search to reduce complexity from $O(N^2)$ to $O(N \\log N)$.
\nTwo sampling strategies were employed:
\nAdditionally, RANSAC inlier thresholds were adjusted per image type based on the 3 classes detected in Stage 1:
\nAfter feature matching, I discovered that careful clustering is essential. Without it, the F1 score (the competition metric) drops significantly. To achieve high-precision clustering, I implemented the following processing.
\n| Component | \nDetails | \n
|---|---|
| Panel Detection | \nYOLOv8-m × 5 (WBF ensemble), trained on synthetic data | \n
| Text Detection | \nYOLOv8-m, manual annotation + synthetic data | \n
| Features | \nSIFT (contrast_threshold=0.001) | \n
| Matching | \nG2NN (α=0.7) + FLANN KDTree | \n
| Geometric Estimation | \nAffine RANSAC + H matrix filtering | \n
| Clustering | \nHDBSCAN + 2-stage merging (H similarity → IoMin) | \n
| Mask Refinement | \nConvex hull + H transformation error-based refinement | \n
Code is available here https://github.com/oshima-yoppi/luc_pub.
", "rawMarkdown": "# Recod.ai/LUC - Scientific Image Forgery Detection 2nd Place Solution\n\nFirst of all, thank you to the organizers for hosting such a wonderful competition. It was a highly meaningful theme—protecting the integrity of scientific research—and I learned a great deal from it.\n\nResult: **Public 3rd → Private 2nd** \n\n## Solution Overview\n\nRather than a deep learning-based approach, my solution is built around **classical feature matching (SIFT + RANSAC)**, combined with YOLO-based preprocessing and custom post-processing in a 3-stage pipeline.\n\n```\nStage 1: Valid Region Detection (Panel/Text detection with YOLOv8)\n ↓\nStage 2: Feature Matching (SIFT + G2NN + RANSAC)\n ↓\nStage 3: Post-processing (Cluster merging + High-precision mask generation)\n```\n---\n\n## Stage 1: Valid Region Detection\n\nIn the subsequent feature matching stage, **graph regions** produce a large number of false positives, and **text regions** cause erroneous matching on caption parts (e.g., matching on the \"mm\" portion of \"10 mm\"). These lead to significant accuracy degradation, making their exclusion essential.\n\nI trained a **Panel detection model** and a **Text detection model** using YOLOv8-m to address this problem.\n\n### Panel Detection\n\nTo exclude irrelevant regions such as graphs, only relevant areas (image regions) are detected. Simultaneously, YOLO classifies detections into 3 classes: **corn images**, **Western blot protein images**, and **other biological images**. The reason for separating classes is that the optimal RANSAC inlier threshold differs by image type (described below).\n\nSince no suitable training data existed, I addressed this by **automatically generating synthetic data**.\n\n- Used a subset of 2,000 images from external data (BioFors dataset) due to time constraints\n- Classified and extracted cell/plant images from the competition's authentic images\n- Auto-generated synthetic graphs (scatter plots, line charts, heatmaps, violin plots)\n- Randomly arranged the above in various layout patterns (1×1, 2×2, 3×1, etc.) to create training data\n- Ensembled 5 YOLOv8-m models trained with different seeds using **Weighted Boxes Fusion (WBF)**\n\n\n### Text Detection\n\nExisting OCR solutions (EasyOCR, etc.) were insufficiently accurate—for example, misrecognizing round cells as the letter \"O\". Since text in scientific paper images has minimal distortion, I determined that OCR-specific heads were unnecessary and **trained YOLOv8-m as an object detector**.\n\n- Manually corrected low-threshold EasyOCR detection results to create annotations (~800 images)\n- Added synthetic data (~2,000 images) with randomly placed text on text-free images\n- Applied diversity in background color, text color, font, rotation, and size for augmentation\n- At inference, performed TTA with two image sizes (640, 1280) and merged both results\n\n---\n\n## Stage 2: Feature Matching\n\n### Aggressive SIFT Keypoint Extraction\n\nBiological images have very little texture, and default settings fail to extract sufficient keypoints.\n\n- Used `contrast_threshold=0.001` (approximately 1/40 of the default 0.04) to extract a large number of keypoints\n- For small images (<1024px), upscaled 4× before feature extraction, then mapped coordinates back to original scale\n\n### G2NN Matching\n\nThe absolute L2 distance values vary greatly across images, making fixed thresholds impractical.\n\nAdopted **G2NN **: sort matching distances and accept only matches up to the breakpoint where $T_{n+1}/T_n > \\alpha$ ($\\alpha=0.7$). This achieves threshold-independent matching.\n\n### Speedup: KDTree\n\nDue to the enormous number of keypoints, used FLANN KDTree approximate nearest neighbor search to reduce complexity from $O(N^2)$ to $O(N \\log N)$.\n\n### RANSAC + Geometric Filtering\n\nTwo sampling strategies were employed:\n\n- **Local sampling**: Execute RANSAC within high-density match regions + YOLO-detected bboxes. Analyze the H matrix to filter out unrealistic transformations (scale >4×, shear deformation, etc.)\n- **Global sampling**: Execute RANSAC once on all matches, then apply HDBSCAN clustering to remove false matches\n\nAdditionally, RANSAC inlier thresholds were adjusted per image type based on the 3 classes detected in Stage 1:\n\n- **Corn images**: Fine black-and-white line patterns cause frequent false matches, so the inlier threshold was set **high** for strict filtering\n- **Protein images (Western blot)**: Very few features make matching inherently difficult, so the inlier threshold was set **low** to enable detection even with few matches\n- **Other biological images**: An intermediate threshold was used\n\n---\n\n## Stage 3: Post-processing\n\nAfter feature matching, I discovered that **careful clustering is essential**. Without it, the F1 score (the competition metric) drops significantly. To achieve high-precision clustering, I implemented the following processing.\n\n### Two-stage Cluster Merging\n\n1. **Stage 1 (within same pair)**: Merge nearby clusters with similar H matrices (normalized difference of rotation angle, scale, and translation < 0.02) using Union-Find\n2. **Stage 2 (across different pairs)**: Merge as the same instance when IoMin (Intersection / min(area1, area2)) between source/dest bboxes ≥ 0.3\n\n### High-precision Mask Generation\n\n1. Recompute H matrix from all corresponding points of merged clusters\n2. Generate source/dest masks via convex hull + dilation\n3. Refine masks precisely using H matrix transformation error (within mean + 2σ)\n4. Apply bidirectional refinement using inverse transformation\n\n---\n\n## What Didn't Work\n\n- **Other feature descriptors** (SURF, ORB, AKAZE, SuperPoint): Could not extract sufficient keypoints in the biological image domain, resulting in poor accuracy\n- **Deep learning-based approaches**: Tried CMFD models combining DINOv2/SegFormer backbones with self-correlation modules, but accuracy was insufficient\n\n---\n\n## Summary\n\n| Component | Details |\n|-----------|---------|\n| Panel Detection | YOLOv8-m × 5 (WBF ensemble), trained on synthetic data |\n| Text Detection | YOLOv8-m, manual annotation + synthetic data |\n| Features | SIFT (contrast_threshold=0.001) |\n| Matching | G2NN (α=0.7) + FLANN KDTree |\n| Geometric Estimation | Affine RANSAC + H matrix filtering |\n| Clustering | HDBSCAN + 2-stage merging (H similarity → IoMin) |\n| Mask Refinement | Convex hull + H transformation error-based refinement |\n\n\nCode is available here https://github.com/oshima-yoppi/luc_pub.", "votes": null }, { "id": "3448686", "postDate": "04/26/2026 11:26:10", "content": "I liked your impressive solution, but the link of your repo doesn't work. Can you fix it ?
", "rawMarkdown": "I liked your impressive solution, but the link of your repo doesn't work. Can you fix it ?", "votes": null }, { "id": "3449075", "postDate": "04/27/2026 12:38:17", "content": "Congrats on the 2nd place, well deserved! 🔥\nVery neat solution and thanks for the writeup!
", "rawMarkdown": "Congrats on the 2nd place, well deserved! 🔥\nVery neat solution and thanks for the writeup!", "votes": null }, { "id": "3449188", "postDate": "04/27/2026 16:37:21", "content": "Thank you! I just made the repository public. The link should work now. Please try again!
", "rawMarkdown": "Thank you! I just made the repository public. The link should work now. Please try again!", "votes": null }, { "id": "3449190", "postDate": "04/27/2026 16:37:58", "content": "thank you very much!!
", "rawMarkdown": "thank you very much!!", "votes": null }, { "id": "3449367", "postDate": "04/28/2026 01:39:51", "content": "Altering SIFT/KDTree/RANSAC parameters by the panel type is a good idea. I also found that SIFT performs poorly on Western blots since there are not that many keypoints. It would be better to have a solution specifically designed for Western Blots.
\nIf I understand correctly, you did not look at any solution which dealt with duplicates inside of the same panel image. I tried forcing it by splitting up every detected panel into halves/thirds/quadrants/sixths. However, it could not complete within the 4 hour runtime. This is especially important for the Western Blot where lanes may be copied over within the same image.
\nBig congrats on 2nd place! I will never look at shiba's the same way again:)
", "rawMarkdown": "Altering SIFT/KDTree/RANSAC parameters by the panel type is a good idea. I also found that SIFT performs poorly on Western blots since there are not that many keypoints. It would be better to have a solution specifically designed for Western Blots.\n\nIf I understand correctly, you did not look at any solution which dealt with duplicates inside of the same panel image. I tried forcing it by splitting up every detected panel into halves/thirds/quadrants/sixths. However, it could not complete within the 4 hour runtime. This is especially important for the Western Blot where lanes may be copied over within the same image.\n\nBig congrats on 2nd place! I will never look at shiba's the same way again:)", "votes": null }, { "id": "3449432", "postDate": "04/28/2026 07:15:07", "content": "Thanks, and again it's a smart solution, well deserved!!
", "rawMarkdown": "Thanks, and again it's a smart solution, well deserved!!", "votes": null }, { "id": "3449587", "postDate": "04/28/2026 15:30:06", "content": "\n\nIt would be better to have a solution specifically designed for Western Blots.
\n
Definitely yes. I struggled a lot with Western blot–type images. It was very difficult to make the method work reliably, and after trying various approaches, I ended up using spatial feature extraction methods such as keypoint-based techniques.
\n\n\nyou did not look at any solution which dealt with duplicates inside of the same panel image
\n
Sorry if I misunderstood your point. In my approach, I am actually able to detect duplicated regions within the same panel. As shown in the example image (which was generated by my algorithm), duplicated regions inside a panel can be successfully identified—even in cases where the entire image is treated as a single panel.\n
How did you manage to do that? I thought from your solution writeup that you only compared SIFT keypoints from 1 panel to the 2nd panel. Or, are the cells you circled \"high density regions\" that you included?
", "rawMarkdown": "How did you manage to do that? I thought from your solution writeup that you only compared SIFT keypoints from 1 panel to the 2nd panel. Or, are the cells you circled \"high density regions\" that you included?", "votes": null }, { "id": "3449846", "postDate": "04/29/2026 00:35:53", "content": "Congrats @npinpi for 2nd Place
", "rawMarkdown": "Congrats @npinpi for 2nd Place", "votes": null }, { "id": "3452910", "postDate": "05/04/2026 14:53:54", "content": "Hello @npinpi and thanks a lot for sharing the solution. I have a question about the stage 1 model, is there any way to access your pretrained models for the panel and text extractors? \nOr like could you be more explicit about how you've used the data to train the models? \nThanks you in advance! cheers
", "rawMarkdown": "Hello @npinpi and thanks a lot for sharing the solution. I have a question about the stage 1 model, is there any way to access your pretrained models for the panel and text extractors? \nOr like could you be more explicit about how you've used the data to train the models? \nThanks you in advance! cheers", "votes": null } ], "comments": [ { "id": 3448686, "author_name": "nacefmbarek", "author_url": "", "post_date": "04/26/2026 11:26:10", "content": "I liked your impressive solution, but the link of your repo doesn't work. Can you fix it ?
", "votes": null, "replies": [ { "id": 3449188, "author_name": "npinpi", "author_url": "", "post_date": "04/27/2026 16:37:21", "content": "Thank you! I just made the repository public. The link should work now. Please try again!
", "votes": null, "replies": [ { "id": 3449432, "author_name": "nacefmbarek", "author_url": "", "post_date": "04/28/2026 07:15:07", "content": "Thanks, and again it's a smart solution, well deserved!!
", "votes": null, "replies": [] }, { "id": 3452910, "author_name": "vasiig", "author_url": "", "post_date": "05/04/2026 14:53:54", "content": "Hello @npinpi and thanks a lot for sharing the solution. I have a question about the stage 1 model, is there any way to access your pretrained models for the panel and text extractors? \nOr like could you be more explicit about how you've used the data to train the models? \nThanks you in advance! cheers
", "votes": null, "replies": [] } ] } ] }, { "id": 3449075, "author_name": "arc144", "author_url": "", "post_date": "04/27/2026 12:38:17", "content": "Congrats on the 2nd place, well deserved! 🔥\nVery neat solution and thanks for the writeup!
", "votes": null, "replies": [ { "id": 3449190, "author_name": "npinpi", "author_url": "", "post_date": "04/27/2026 16:37:58", "content": "thank you very much!!
", "votes": null, "replies": [] } ] }, { "id": 3449367, "author_name": "returnofsputnik", "author_url": "", "post_date": "04/28/2026 01:39:51", "content": "Altering SIFT/KDTree/RANSAC parameters by the panel type is a good idea. I also found that SIFT performs poorly on Western blots since there are not that many keypoints. It would be better to have a solution specifically designed for Western Blots.
\nIf I understand correctly, you did not look at any solution which dealt with duplicates inside of the same panel image. I tried forcing it by splitting up every detected panel into halves/thirds/quadrants/sixths. However, it could not complete within the 4 hour runtime. This is especially important for the Western Blot where lanes may be copied over within the same image.
\nBig congrats on 2nd place! I will never look at shiba's the same way again:)
", "votes": null, "replies": [ { "id": 3449587, "author_name": "npinpi", "author_url": "", "post_date": "04/28/2026 15:30:06", "content": "\n\nIt would be better to have a solution specifically designed for Western Blots.
\n
Definitely yes. I struggled a lot with Western blot–type images. It was very difficult to make the method work reliably, and after trying various approaches, I ended up using spatial feature extraction methods such as keypoint-based techniques.
\n\n\nyou did not look at any solution which dealt with duplicates inside of the same panel image
\n
Sorry if I misunderstood your point. In my approach, I am actually able to detect duplicated regions within the same panel. As shown in the example image (which was generated by my algorithm), duplicated regions inside a panel can be successfully identified—even in cases where the entire image is treated as a single panel.\n
How did you manage to do that? I thought from your solution writeup that you only compared SIFT keypoints from 1 panel to the 2nd panel. Or, are the cells you circled \"high density regions\" that you included?
", "votes": null, "replies": [] } ] } ] }, { "id": 3449846, "author_name": "navneetbende", "author_url": "", "post_date": "04/29/2026 00:35:53", "content": "Congrats @npinpi for 2nd Place
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3448118": "# Recod.ai/LUC - Scientific Image Forgery Detection 2nd Place Solution\n\nFirst of all, thank you to the organizers for hosting such a wonderful competition. It was a highly meaningful theme—protecting the integrity of scientific research—and I learned a great deal from it.\n\nResult: **Public 3rd → Private 2nd** \n\n## Solution Overview\n\nRather than a deep learning-based approach, my solution is built around **classical feature matching (SIFT + RANSAC)**, combined with YOLO-based preprocessing and custom post-processing in a 3-stage pipeline.\n\n```\nStage 1: Valid Region Detection (Panel/Text detection with YOLOv8)\n ↓\nStage 2: Feature Matching (SIFT + G2NN + RANSAC)\n ↓\nStage 3: Post-processing (Cluster merging + High-precision mask generation)\n```\n---\n\n## Stage 1: Valid Region Detection\n\nIn the subsequent feature matching stage, **graph regions** produce a large number of false positives, and **text regions** cause erroneous matching on caption parts (e.g., matching on the \"mm\" portion of \"10 mm\"). These lead to significant accuracy degradation, making their exclusion essential.\n\nI trained a **Panel detection model** and a **Text detection model** using YOLOv8-m to address this problem.\n\n### Panel Detection\n\nTo exclude irrelevant regions such as graphs, only relevant areas (image regions) are detected. Simultaneously, YOLO classifies detections into 3 classes: **corn images**, **Western blot protein images**, and **other biological images**. The reason for separating classes is that the optimal RANSAC inlier threshold differs by image type (described below).\n\nSince no suitable training data existed, I addressed this by **automatically generating synthetic data**.\n\n- Used a subset of 2,000 images from external data (BioFors dataset) due to time constraints\n- Classified and extracted cell/plant images from the competition's authentic images\n- Auto-generated synthetic graphs (scatter plots, line charts, heatmaps, violin plots)\n- Randomly arranged the above in various layout patterns (1×1, 2×2, 3×1, etc.) to create training data\n- Ensembled 5 YOLOv8-m models trained with different seeds using **Weighted Boxes Fusion (WBF)**\n\n\n### Text Detection\n\nExisting OCR solutions (EasyOCR, etc.) were insufficiently accurate—for example, misrecognizing round cells as the letter \"O\". Since text in scientific paper images has minimal distortion, I determined that OCR-specific heads were unnecessary and **trained YOLOv8-m as an object detector**.\n\n- Manually corrected low-threshold EasyOCR detection results to create annotations (~800 images)\n- Added synthetic data (~2,000 images) with randomly placed text on text-free images\n- Applied diversity in background color, text color, font, rotation, and size for augmentation\n- At inference, performed TTA with two image sizes (640, 1280) and merged both results\n\n---\n\n## Stage 2: Feature Matching\n\n### Aggressive SIFT Keypoint Extraction\n\nBiological images have very little texture, and default settings fail to extract sufficient keypoints.\n\n- Used `contrast_threshold=0.001` (approximately 1/40 of the default 0.04) to extract a large number of keypoints\n- For small images (<1024px), upscaled 4× before feature extraction, then mapped coordinates back to original scale\n\n### G2NN Matching\n\nThe absolute L2 distance values vary greatly across images, making fixed thresholds impractical.\n\nAdopted **G2NN **: sort matching distances and accept only matches up to the breakpoint where $T_{n+1}/T_n > \\alpha$ ($\\alpha=0.7$). This achieves threshold-independent matching.\n\n### Speedup: KDTree\n\nDue to the enormous number of keypoints, used FLANN KDTree approximate nearest neighbor search to reduce complexity from $O(N^2)$ to $O(N \\log N)$.\n\n### RANSAC + Geometric Filtering\n\nTwo sampling strategies were employed:\n\n- **Local sampling**: Execute RANSAC within high-density match regions + YOLO-detected bboxes. Analyze the H matrix to filter out unrealistic transformations (scale >4×, shear deformation, etc.)\n- **Global sampling**: Execute RANSAC once on all matches, then apply HDBSCAN clustering to remove false matches\n\nAdditionally, RANSAC inlier thresholds were adjusted per image type based on the 3 classes detected in Stage 1:\n\n- **Corn images**: Fine black-and-white line patterns cause frequent false matches, so the inlier threshold was set **high** for strict filtering\n- **Protein images (Western blot)**: Very few features make matching inherently difficult, so the inlier threshold was set **low** to enable detection even with few matches\n- **Other biological images**: An intermediate threshold was used\n\n---\n\n## Stage 3: Post-processing\n\nAfter feature matching, I discovered that **careful clustering is essential**. Without it, the F1 score (the competition metric) drops significantly. To achieve high-precision clustering, I implemented the following processing.\n\n### Two-stage Cluster Merging\n\n1. **Stage 1 (within same pair)**: Merge nearby clusters with similar H matrices (normalized difference of rotation angle, scale, and translation < 0.02) using Union-Find\n2. **Stage 2 (across different pairs)**: Merge as the same instance when IoMin (Intersection / min(area1, area2)) between source/dest bboxes ≥ 0.3\n\n### High-precision Mask Generation\n\n1. Recompute H matrix from all corresponding points of merged clusters\n2. Generate source/dest masks via convex hull + dilation\n3. Refine masks precisely using H matrix transformation error (within mean + 2σ)\n4. Apply bidirectional refinement using inverse transformation\n\n---\n\n## What Didn't Work\n\n- **Other feature descriptors** (SURF, ORB, AKAZE, SuperPoint): Could not extract sufficient keypoints in the biological image domain, resulting in poor accuracy\n- **Deep learning-based approaches**: Tried CMFD models combining DINOv2/SegFormer backbones with self-correlation modules, but accuracy was insufficient\n\n---\n\n## Summary\n\n| Component | Details |\n|-----------|---------|\n| Panel Detection | YOLOv8-m × 5 (WBF ensemble), trained on synthetic data |\n| Text Detection | YOLOv8-m, manual annotation + synthetic data |\n| Features | SIFT (contrast_threshold=0.001) |\n| Matching | G2NN (α=0.7) + FLANN KDTree |\n| Geometric Estimation | Affine RANSAC + H matrix filtering |\n| Clustering | HDBSCAN + 2-stage merging (H similarity → IoMin) |\n| Mask Refinement | Convex hull + H transformation error-based refinement |\n\n\nCode is available here https://github.com/oshima-yoppi/luc_pub.", "3448686": "I liked your impressive solution, but the link of your repo doesn't work. Can you fix it ?", "3449075": "Congrats on the 2nd place, well deserved! 🔥\nVery neat solution and thanks for the writeup!", "3449188": "Thank you! I just made the repository public. The link should work now. Please try again!", "3449190": "thank you very much!!", "3449367": "Altering SIFT/KDTree/RANSAC parameters by the panel type is a good idea. I also found that SIFT performs poorly on Western blots since there are not that many keypoints. It would be better to have a solution specifically designed for Western Blots.\n\nIf I understand correctly, you did not look at any solution which dealt with duplicates inside of the same panel image. I tried forcing it by splitting up every detected panel into halves/thirds/quadrants/sixths. However, it could not complete within the 4 hour runtime. This is especially important for the Western Blot where lanes may be copied over within the same image.\n\nBig congrats on 2nd place! I will never look at shiba's the same way again:)", "3449432": "Thanks, and again it's a smart solution, well deserved!!", "3449587": "> It would be better to have a solution specifically designed for Western Blots.\n\nDefinitely yes. I struggled a lot with Western blot–type images. It was very difficult to make the method work reliably, and after trying various approaches, I ended up using spatial feature extraction methods such as keypoint-based techniques.\n\n> you did not look at any solution which dealt with duplicates inside of the same panel image\n\nSorry if I misunderstood your point. In my approach, I am actually able to detect duplicated regions within the same panel. As shown in the example image (which was generated by my algorithm), duplicated regions inside a panel can be successfully identified—even in cases where the entire image is treated as a single panel.\n", "3449797": "How did you manage to do that? I thought from your solution writeup that you only compared SIFT keypoints from 1 panel to the 2nd panel. Or, are the cells you circled \"high density regions\" that you included?", "3449846": "Congrats @npinpi for 2nd Place", "3452910": "Hello @npinpi and thanks a lot for sharing the solution. I have a question about the stage 1 model, is there any way to access your pretrained models for the panel and text extractors? \nOr like could you be more explicit about how you've used the data to train the models? \nThanks you in advance! cheers" }, "source": "meta" }