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 (out of 1564 teams)
\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)\nThe overall algorithm overview is shown in the image below.
\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 (Good-to-Next Neighbor): 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 $O(N \\log N)$.
\nHere, matching features are detected based on keypoints within the valid regions detected in Stage 1. Two sampling strategies were employed for detection:
\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
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 (out of 1564 teams)
\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)\nThe overall algorithm overview is shown in the image below.
\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 (Good-to-Next Neighbor): 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 $O(N \\log N)$.
\nHere, matching features are detected based on keypoints within the valid regions detected in Stage 1. Two sampling strategies were employed for detection:
\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