As a great distraction from the final hours of the LB I like to remember the real point of our efforts here on Kaggle. Not to chase some LB but to learn more skills! Please feel free to share whatever you thought you learned too! Good luck and happy Kaggling!
\nYOLO Object Detection Basics - Its no secret YOLO was very commonly used for this competition, but there was extra need to get familiar with how to change augments, what was available, changing training parameters in useful ways, and much more. I havent used YOLO very much and had room for grown in my object detection skills, as typically I would have wanted to do 2.5d modeling instead.
YOLO Ensembling - The blending of these models ended up being surprisingly unique. I wont share too much on this so early but some good information somewhat similar can be found here: https://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/578156
Medical imaging specifics - Learning the way these images were marked to the best of our ability is a huge difference in this comp. Understanding extra data, voxel spacing, the way tomos are separated, and more are super important to output. I assume that a few at the absolute top of the LB have found a little extra in this regard.
Napari - A great tool for doing a lot of these object detection markings as many labeled their own data. My theory is that this is not only helpful to get extra data but it also helps us with our true task of trying to predict the labeler which seems to be the real subtask behind/besides the data itself.
Distributed compute - For 5 subs a day which was all but a requirement for this competition training speed mattered a lot so being able to 1 use multi-GPU for YOLO and 2 multi-GPU for the script was important to be able to train 5 models a day, especially using extra data. I had a decent grasp of this but doing this with YOLO for whatever reason was quite a bit different.
Extra 2.5d Modeling - I originally really liked the idea of 2.5d as this was common in my previous domain of medical imaging but for this challenge it wasnt extremely helpful. I am assuming that this is something to do with the spacing and the way the data is given.
Very random domain knowledge - Tons of interesting knowledge about this field specifically as our team, did tons of research to be able to try to get any edge that may help us. We thought that this helped us with dataset creation but LB changes and basic shift between train and test along with sample size being so small made it extremely difficult to paint a full picture of what we are trying to solve for. I would think that without the domain shift this would be an effectively \"Solvable\" problem? As can be seen that you can get an internal train and val of 98%+ very easily.
Good luck!
", "messages": [ { "id": 3217054, "postDate": "2025-06-04T13:26:15.817Z", "content": "As a great distraction from the final hours of the LB I like to remember the real point of our efforts here on Kaggle. Not to chase some LB but to learn more skills! Please feel free to share whatever you thought you learned too! Good luck and happy Kaggling!
\nYOLO Object Detection Basics - Its no secret YOLO was very commonly used for this competition, but there was extra need to get familiar with how to change augments, what was available, changing training parameters in useful ways, and much more. I havent used YOLO very much and had room for grown in my object detection skills, as typically I would have wanted to do 2.5d modeling instead.
YOLO Ensembling - The blending of these models ended up being surprisingly unique. I wont share too much on this so early but some good information somewhat similar can be found here: https://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/578156
Medical imaging specifics - Learning the way these images were marked to the best of our ability is a huge difference in this comp. Understanding extra data, voxel spacing, the way tomos are separated, and more are super important to output. I assume that a few at the absolute top of the LB have found a little extra in this regard.
Napari - A great tool for doing a lot of these object detection markings as many labeled their own data. My theory is that this is not only helpful to get extra data but it also helps us with our true task of trying to predict the labeler which seems to be the real subtask behind/besides the data itself.
Distributed compute - For 5 subs a day which was all but a requirement for this competition training speed mattered a lot so being able to 1 use multi-GPU for YOLO and 2 multi-GPU for the script was important to be able to train 5 models a day, especially using extra data. I had a decent grasp of this but doing this with YOLO for whatever reason was quite a bit different.
Extra 2.5d Modeling - I originally really liked the idea of 2.5d as this was common in my previous domain of medical imaging but for this challenge it wasnt extremely helpful. I am assuming that this is something to do with the spacing and the way the data is given.
Very random domain knowledge - Tons of interesting knowledge about this field specifically as our team, did tons of research to be able to try to get any edge that may help us. We thought that this helped us with dataset creation but LB changes and basic shift between train and test along with sample size being so small made it extremely difficult to paint a full picture of what we are trying to solve for. I would think that without the domain shift this would be an effectively \"Solvable\" problem? As can be seen that you can get an internal train and val of 98%+ very easily.
Good luck!
", "rawMarkdown": "As a great distraction from the final hours of the LB I like to remember the real point of our efforts here on Kaggle. Not to chase some LB but to learn more skills! Please feel free to share whatever you thought you learned too! Good luck and happy Kaggling!\n\n1. YOLO Object Detection Basics - Its no secret YOLO was very commonly used for this competition, but there was extra need to get familiar with how to change augments, what was available, changing training parameters in useful ways, and much more. I havent used YOLO very much and had room for grown in my object detection skills, as typically I would have wanted to do 2.5d modeling instead. \n\n2. YOLO Ensembling - The blending of these models ended up being surprisingly unique. I wont share too much on this so early but some good information somewhat similar can be found here: https://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/578156\n\n3. Medical imaging specifics - Learning the way these images were marked to the best of our ability is a huge difference in this comp. Understanding extra data, voxel spacing, the way tomos are separated, and more are super important to output. I assume that a few at the absolute top of the LB have found a little extra in this regard. \n\n4. Napari - A great tool for doing a lot of these object detection markings as many labeled their own data. My theory is that this is not only helpful to get extra data but it also helps us with our true task of trying to predict the labeler which seems to be the real subtask behind/besides the data itself. \n\n5. Distributed compute - For 5 subs a day which was all but a requirement for this competition training speed mattered a lot so being able to 1 use multi-GPU for YOLO and 2 multi-GPU for the script was important to be able to train 5 models a day, especially using extra data. I had a decent grasp of this but doing this with YOLO for whatever reason was quite a bit different. \n\n6. Extra 2.5d Modeling - I originally really liked the idea of 2.5d as this was common in my previous domain of medical imaging but for this challenge it wasnt extremely helpful. I am assuming that this is something to do with the spacing and the way the data is given. \n\n7. Very random domain knowledge - Tons of interesting knowledge about this field specifically as our team, did tons of research to be able to try to get any edge that may help us. We thought that this helped us with dataset creation but LB changes and basic shift between train and test along with sample size being so small made it extremely difficult to paint a full picture of what we are trying to solve for. I would think that without the domain shift this would be an effectively \"Solvable\" problem? As can be seen that you can get an internal train and val of 98%+ very easily. \n\nGood luck! ", "votes": 26 }, { "id": 3217092, "postDate": "2025-06-04T14:18:35.890Z", "content": "I'm very happy to know @cody11null and @andrewjdarley. You both are amazing people. And really thank to Andrew for hosting such good competition, it helps a lot for my job interview. Look forward to future cooperation!
", "rawMarkdown": "I'm very happy to know @cody11null and @andrewjdarley. You both are amazing people. And really thank to Andrew for hosting such good competition, it helps a lot for my job interview. Look forward to future cooperation!", "votes": 3, "replies": [ { "id": 3217102, "postDate": "2025-06-04T14:35:17.647Z", "content": "I am looking forward to your solution @tom99763 , you really had very nice ideas and insights in this competition!
", "rawMarkdown": "I am looking forward to your solution @tom99763 , you really had very nice ideas and insights in this competition!", "votes": 3, "replies": [ { "id": 3217126, "postDate": "2025-06-04T14:57:15.470Z", "content": "I think we all are for this one! I just hope it holds through the shake to get the medal that is deserved!
", "rawMarkdown": "I think we all are for this one! I just hope it holds through the shake to get the medal that is deserved! ", "votes": 2 }, { "id": 3217181, "postDate": "2025-06-04T16:01:34.437Z", "content": "I think I cannot complete write-up tonight… tomorrow I'll post the diagram first then progressively fill the content!
", "rawMarkdown": "I think I cannot complete write-up tonight... tomorrow I'll post the diagram first then progressively fill the content!", "votes": 1 } ] } ] }, { "id": 3217160, "postDate": "2025-06-04T15:34:46.280Z", "content": "The YOLO algorithm is indeed excellent, but it has a drawback in that its recall rate is relatively low, and it often misses detecting some targets.
", "rawMarkdown": "The YOLO algorithm is indeed excellent, but it has a drawback in that its recall rate is relatively low, and it often misses detecting some targets.", "votes": 4 }, { "id": 3217156, "postDate": "2025-06-04T15:28:11.657Z", "content": "I didn’t do much this time apart from exploring different YOLO models. I wasn’t able to ensemble well either, so the competition left me feeling a bit disappointed. External factors threw me off a bit too, which made it harder to stay focused.
", "rawMarkdown": "I didn’t do much this time apart from exploring different YOLO models. I wasn’t able to ensemble well either, so the competition left me feeling a bit disappointed. External factors threw me off a bit too, which made it harder to stay focused.", "votes": 1 }, { "id": 3217074, "postDate": "2025-06-04T13:52:58.530Z", "content": "And! Thank you @andrewjdarley for being a very interactive/responsive host!
", "rawMarkdown": "And! Thank you @andrewjdarley for being a very interactive/responsive host!", "votes": 2 }, { "id": 3217342, "postDate": "2025-06-04T23:08:07.400Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3217092, "author_name": "Tom", "author_url": "", "post_date": "2025-06-04T14:18:35.890000", "content": "I'm very happy to know @cody11null and @andrewjdarley. You both are amazing people. And really thank to Andrew for hosting such good competition, it helps a lot for my job interview. Look forward to future cooperation!
", "votes": 3, "replies": [ { "id": 3217102, "author_name": "IAmParadox", "author_url": "", "post_date": "2025-06-04T14:35:17.647000", "content": "I am looking forward to your solution @tom99763 , you really had very nice ideas and insights in this competition!
", "votes": 3, "replies": [ { "id": 3217126, "author_name": "Cody_Null", "author_url": "", "post_date": "2025-06-04T14:57:15.470000", "content": "I think we all are for this one! I just hope it holds through the shake to get the medal that is deserved!
", "votes": 2, "replies": [] }, { "id": 3217181, "author_name": "Tom", "author_url": "", "post_date": "2025-06-04T16:01:34.437000", "content": "I think I cannot complete write-up tonight… tomorrow I'll post the diagram first then progressively fill the content!
", "votes": 1, "replies": [] } ] } ] }, { "id": 3217160, "author_name": "yyyy0201", "author_url": "", "post_date": "2025-06-04T15:34:46.280000", "content": "The YOLO algorithm is indeed excellent, but it has a drawback in that its recall rate is relatively low, and it often misses detecting some targets.
", "votes": 4, "replies": [] }, { "id": 3217156, "author_name": "CodeHacker", "author_url": "", "post_date": "2025-06-04T15:28:11.657000", "content": "I didn’t do much this time apart from exploring different YOLO models. I wasn’t able to ensemble well either, so the competition left me feeling a bit disappointed. External factors threw me off a bit too, which made it harder to stay focused.
", "votes": 1, "replies": [] }, { "id": 3217074, "author_name": "Cody_Null", "author_url": "", "post_date": "2025-06-04T13:52:58.530000", "content": "And! Thank you @andrewjdarley for being a very interactive/responsive host!
", "votes": 2, "replies": [] }, { "id": 3217342, "author_name": "", "author_url": "", "post_date": "2025-06-04T23:08:07.400000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3217054": "As a great distraction from the final hours of the LB I like to remember the real point of our efforts here on Kaggle. Not to chase some LB but to learn more skills! Please feel free to share whatever you thought you learned too! Good luck and happy Kaggling!\n\n1. YOLO Object Detection Basics - Its no secret YOLO was very commonly used for this competition, but there was extra need to get familiar with how to change augments, what was available, changing training parameters in useful ways, and much more. I havent used YOLO very much and had room for grown in my object detection skills, as typically I would have wanted to do 2.5d modeling instead. \n\n2. YOLO Ensembling - The blending of these models ended up being surprisingly unique. I wont share too much on this so early but some good information somewhat similar can be found here: https://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/578156\n\n3. Medical imaging specifics - Learning the way these images were marked to the best of our ability is a huge difference in this comp. Understanding extra data, voxel spacing, the way tomos are separated, and more are super important to output. I assume that a few at the absolute top of the LB have found a little extra in this regard. \n\n4. Napari - A great tool for doing a lot of these object detection markings as many labeled their own data. My theory is that this is not only helpful to get extra data but it also helps us with our true task of trying to predict the labeler which seems to be the real subtask behind/besides the data itself. \n\n5. Distributed compute - For 5 subs a day which was all but a requirement for this competition training speed mattered a lot so being able to 1 use multi-GPU for YOLO and 2 multi-GPU for the script was important to be able to train 5 models a day, especially using extra data. I had a decent grasp of this but doing this with YOLO for whatever reason was quite a bit different. \n\n6. Extra 2.5d Modeling - I originally really liked the idea of 2.5d as this was common in my previous domain of medical imaging but for this challenge it wasnt extremely helpful. I am assuming that this is something to do with the spacing and the way the data is given. \n\n7. Very random domain knowledge - Tons of interesting knowledge about this field specifically as our team, did tons of research to be able to try to get any edge that may help us. We thought that this helped us with dataset creation but LB changes and basic shift between train and test along with sample size being so small made it extremely difficult to paint a full picture of what we are trying to solve for. I would think that without the domain shift this would be an effectively \"Solvable\" problem? As can be seen that you can get an internal train and val of 98%+ very easily. \n\nGood luck! ", "3217092": "I'm very happy to know @cody11null and @andrewjdarley. You both are amazing people. And really thank to Andrew for hosting such good competition, it helps a lot for my job interview. Look forward to future cooperation!", "3217160": "The YOLO algorithm is indeed excellent, but it has a drawback in that its recall rate is relatively low, and it often misses detecting some targets.", "3217156": "I didn’t do much this time apart from exploring different YOLO models. I wasn’t able to ensemble well either, so the competition left me feeling a bit disappointed. External factors threw me off a bit too, which made it harder to stay focused.", "3217074": "And! Thank you @andrewjdarley for being a very interactive/responsive host!", "3217342": "" } }