First of all, I would like to thank @playwithme for sharing an excellent notebook. This is the notebook I referred to during the competition.
\nLB0.81-MHAFYOLO -tta-wbf-Submission Notebook(https://www.kaggle.com/code/playwithme/lb0-81-mhafyolo-tta-wbf-submission-notebook)
\nThe model used in the notebook (https://www.kaggle.com/models/yyyy0201/mhafyolo/PyTorch/default) was pre-trained and publicly available.
My main improvement was quite simple — I only modified one of the functions
\n ():\n \n\n detections:\n []\n\n \n xy_distance_threshold = * iou_threshold \n z_distance_threshold = \n min_z_span = \n\n \n detections = (detections, key= d: d[], reverse=)\n final_detections = []\n\n detections:\n base = detections.pop()\n group = [base]\n z_set = {base[]}\n rest = []\n\n d detections:\n dz = (d[] - base[])\n dy = (d[] - base[])\n dx = (d[] - base[])\n dz <= z_distance_threshold dy <= xy_distance_threshold dx <= xy_distance_threshold:\n group.append(d)\n z_set.add(d[])\n :\n rest.append(d)\n\n \n (z_set) >= min_z_span:\n best = (group, key= g: g[])\n final_detections.append(best)\n\n detections = rest\n\n final_detections\nFunction Overview
\nThe perform_3d_nms function performs 3D clustering of detection points, mimicking a Non-Maximum Suppression (NMS) mechanism across 3D space. It filters out redundant detections by identifying spatially close points across Z-slices and keeps only the most confident detection within each valid cluster.
Parameters
\n• detections: A list of detection dictionaries, each with keys 'z', 'y', 'x', and 'confidence'.
\n• iou_threshold: A scaling factor for the XY distance threshold (e.g., 0.2).
How It Works
\nReturns
\n• A list of representative detections (with the highest confidence) from valid 3D clusters.
The following hyperparameters were used.
\nCONFIDENCE_THRESHOLD = 0.7 # Lower threshold to catch more potential motors
\nMAX_DETECTIONS_PER_TOMO = 1 # Keep track of top N detections per tomogram
\nNMS_IOU_THRESHOLD = 0.7 # Non-maximum suppression threshold for 3D clustering
\nCONCENTRATION = 1 # ONLY PROCESS 1/20 slices for fast submission
\nSIZE = 1024
First of all, I would like to thank @playwithme for sharing an excellent notebook. This is the notebook I referred to during the competition.
\nLB0.81-MHAFYOLO -tta-wbf-Submission Notebook(https://www.kaggle.com/code/playwithme/lb0-81-mhafyolo-tta-wbf-submission-notebook)
\nThe model used in the notebook (https://www.kaggle.com/models/yyyy0201/mhafyolo/PyTorch/default) was pre-trained and publicly available.
My main improvement was quite simple — I only modified one of the functions
\n ():\n \n\n detections:\n []\n\n \n xy_distance_threshold = * iou_threshold \n z_distance_threshold = \n min_z_span = \n\n \n detections = (detections, key= d: d[], reverse=)\n final_detections = []\n\n detections:\n base = detections.pop()\n group = [base]\n z_set = {base[]}\n rest = []\n\n d detections:\n dz = (d[] - base[])\n dy = (d[] - base[])\n dx = (d[] - base[])\n dz <= z_distance_threshold dy <= xy_distance_threshold dx <= xy_distance_threshold:\n group.append(d)\n z_set.add(d[])\n :\n rest.append(d)\n\n \n (z_set) >= min_z_span:\n best = (group, key= g: g[])\n final_detections.append(best)\n\n detections = rest\n\n final_detections\nFunction Overview
\nThe perform_3d_nms function performs 3D clustering of detection points, mimicking a Non-Maximum Suppression (NMS) mechanism across 3D space. It filters out redundant detections by identifying spatially close points across Z-slices and keeps only the most confident detection within each valid cluster.
Parameters
\n• detections: A list of detection dictionaries, each with keys 'z', 'y', 'x', and 'confidence'.
\n• iou_threshold: A scaling factor for the XY distance threshold (e.g., 0.2).
How It Works
\nReturns
\n• A list of representative detections (with the highest confidence) from valid 3D clusters.
The following hyperparameters were used.
\nCONFIDENCE_THRESHOLD = 0.7 # Lower threshold to catch more potential motors
\nMAX_DETECTIONS_PER_TOMO = 1 # Keep track of top N detections per tomogram
\nNMS_IOU_THRESHOLD = 0.7 # Non-maximum suppression threshold for 3D clustering
\nCONCENTRATION = 1 # ONLY PROCESS 1/20 slices for fast submission
\nSIZE = 1024
Nice solution, we got halfway through RTDETRv2 not as good as my own YOLO model😭
", "rawMarkdown": "Nice solution, we got halfway through RTDETRv2 not as good as my own YOLO model😭", "votes": 1, "replies": [ { "id": 3218426, "postDate": "2025-06-06T07:16:05.827Z", "content": "show your Solution! 😆
", "rawMarkdown": "show your Solution! 😆" } ] }, { "id": 3218270, "postDate": "2025-06-06T03:36:21.203Z", "content": "Good job on te solution 🤗
", "rawMarkdown": "Good job on te solution 🤗", "votes": 1 }, { "id": 3218202, "postDate": "2025-06-06T01:14:43.283Z", "content": "Well Done, but I ended up not submitting with this weight in the end, hahaha.😭😭😭
", "rawMarkdown": "Well Done, but I ended up not submitting with this weight in the end, hahaha.😭😭😭", "votes": 1 }, { "id": 3218193, "postDate": "2025-06-06T00:54:13.793Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3218413, "author_name": "yyyy0201", "author_url": "", "post_date": "2025-06-06T06:39:26.137000", "content": "Nice solution, we got halfway through RTDETRv2 not as good as my own YOLO model😭
", "votes": 1, "replies": [ { "id": 3218426, "author_name": "Hide on bush", "author_url": "", "post_date": "2025-06-06T07:16:05.827000", "content": "show your Solution! 😆
", "votes": 0, "replies": [] } ] }, { "id": 3218270, "author_name": "Joseph Hadad", "author_url": "", "post_date": "2025-06-06T03:36:21.203000", "content": "Good job on te solution 🤗
", "votes": 1, "replies": [] }, { "id": 3218202, "author_name": "Hide on bush", "author_url": "", "post_date": "2025-06-06T01:14:43.283000", "content": "Well Done, but I ended up not submitting with this weight in the end, hahaha.😭😭😭
", "votes": 1, "replies": [] }, { "id": 3218193, "author_name": "", "author_url": "", "post_date": "2025-06-06T00:54:13.793000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3218186": "First of all, I would like to thank @playwithme for sharing an excellent notebook. This is the notebook I referred to during the competition.\nLB0.81-MHAFYOLO -tta-wbf-Submission Notebook(https://www.kaggle.com/code/playwithme/lb0-81-mhafyolo-tta-wbf-submission-notebook)\nThe model used in the notebook (https://www.kaggle.com/models/yyyy0201/mhafyolo/PyTorch/default) was pre-trained and publicly available.\n\nMy main improvement was quite simple — I only modified one of the functions\n\n```python\ndef perform_3d_nms(detections, iou_threshold):\n \"\"\"\n 3D NMS-like clustering that considers spatial distance and Z-slice thickness.\n\n Parameters:\n detections (list of dict): Detection info list [{'z': int, 'y': int, 'x': int, 'confidence': float}, ...]\n iou_threshold (float): Threshold to scale the XY spatial distance limit (e.g., 0.2)\n\n Returns:\n list of dict: Representative detections after NMS (can include multiple clusters)\n \"\"\"\n\n if not detections:\n return []\n\n # Clustering conditions\n xy_distance_threshold = 24 * iou_threshold # 24 is the assumed side length of a box\n z_distance_threshold = 10 # Acceptable range in Z-slice direction\n min_z_span = 3 # Minimum number of Z-slices to consider as a valid cluster\n\n # Sort detections by confidence in descending order\n detections = sorted(detections, key=lambda d: d['confidence'], reverse=True)\n final_detections = []\n\n while detections:\n base = detections.pop(0)\n group = [base]\n z_set = {base['z']}\n rest = []\n\n for d in detections:\n dz = abs(d['z'] - base['z'])\n dy = abs(d['y'] - base['y'])\n dx = abs(d['x'] - base['x'])\n if dz <= z_distance_threshold and dy <= xy_distance_threshold and dx <= xy_distance_threshold:\n group.append(d)\n z_set.add(d['z'])\n else:\n rest.append(d)\n\n # If the group spans enough Z-slices, take the highest-confidence detection\n if len(z_set) >= min_z_span:\n best = max(group, key=lambda g: g['confidence'])\n final_detections.append(best)\n\n detections = rest\n\n return final_detections\n\n```\n\nFunction Overview\nThe perform_3d_nms function performs 3D clustering of detection points, mimicking a Non-Maximum Suppression (NMS) mechanism across 3D space. It filters out redundant detections by identifying spatially close points across Z-slices and keeps only the most confident detection within each valid cluster.\n________________________________________\nParameters\n•\tdetections: A list of detection dictionaries, each with keys 'z', 'y', 'x', and 'confidence'.\n•\tiou_threshold: A scaling factor for the XY distance threshold (e.g., 0.2).\n________________________________________\nHow It Works\n1.\tThreshold Definition:\no\tXY distance threshold = 24 * iou_threshold\no\tZ-slice range threshold = 10\no\tMinimum Z-slice span to be a valid cluster = 3\n2.\tSort detections by confidence (descending)\n3.\tClustering Process:\no\tUse the most confident detection as the base.\no\tGroup together detections that are close in Z, Y, and X directions.\no\tIf the group spans at least 3 Z-slices, keep the one with the highest confidence.\n4.\tRemove processed detections and repeat\n________________________________________\nReturns\n•\tA list of representative detections (with the highest confidence) from valid 3D clusters.\n\nThe following hyperparameters were used.\nCONFIDENCE_THRESHOLD = 0.7 # Lower threshold to catch more potential motors\nMAX_DETECTIONS_PER_TOMO = 1 # Keep track of top N detections per tomogram\nNMS_IOU_THRESHOLD = 0.7 # Non-maximum suppression threshold for 3D clustering\nCONCENTRATION = 1 # ONLY PROCESS 1/20 slices for fast submission\nSIZE = 1024\n", "3218413": "Nice solution, we got halfway through RTDETRv2 not as good as my own YOLO model😭", "3218270": "Good job on te solution 🤗", "3218202": "Well Done, but I ended up not submitting with this weight in the end, hahaha.😭😭😭", "3218193": "" } }