Thanks to Kaggle and the organizers of the competition for the baseline provided. This is our first time participating in an image object detection competition, and we have learned a lot.
\nWe used different random seeds, divided the training and validation sets multiple times, and experimented with various image sizes (640, 840, 960, 1280) and YOLO versions. We also adopted several data augmentation techniques, such as mosaic, mixup, perspective, shear, and auto augmentation.
\nWe found that the four models achieved the highest LB score when integrated. Our self-trained YOLO v8 and YOLO v10 models, combined with YOLO v10 and YOLO v11 models from public datasets, yielded the highest scores. We are fortunate to have chosen the combination that achieved the highest private score.
\nFinally, thanks to YukiZ and Eng Adam Al Mohammedi for making the code and model public.
", "messages": [ { "id": 3217534, "postDate": "2025-06-05T05:41:48.517Z", "content": "Thanks to Kaggle and the organizers of the competition for the baseline provided. This is our first time participating in an image object detection competition, and we have learned a lot.
\nWe used different random seeds, divided the training and validation sets multiple times, and experimented with various image sizes (640, 840, 960, 1280) and YOLO versions. We also adopted several data augmentation techniques, such as mosaic, mixup, perspective, shear, and auto augmentation.
\nWe found that the four models achieved the highest LB score when integrated. Our self-trained YOLO v8 and YOLO v10 models, combined with YOLO v10 and YOLO v11 models from public datasets, yielded the highest scores. We are fortunate to have chosen the combination that achieved the highest private score.
\nFinally, thanks to YukiZ and Eng Adam Al Mohammedi for making the code and model public.
", "rawMarkdown": "Thanks to Kaggle and the organizers of the competition for the baseline provided. This is our first time participating in an image object detection competition, and we have learned a lot.\n\nWe used different random seeds, divided the training and validation sets multiple times, and experimented with various image sizes (640, 840, 960, 1280) and YOLO versions. We also adopted several data augmentation techniques, such as mosaic, mixup, perspective, shear, and auto augmentation.\n\nWe found that the four models achieved the highest LB score when integrated. Our self-trained YOLO v8 and YOLO v10 models, combined with YOLO v10 and YOLO v11 models from public datasets, yielded the highest scores. We are fortunate to have chosen the combination that achieved the highest private score.\n\nFinally, thanks to YukiZ and Eng Adam Al Mohammedi for making the code and model public.", "votes": 5 }, { "id": 3217813, "postDate": "2025-06-05T12:27:52.627Z", "content": "Hello! Can you show your inference code? Because I'm trying to integrate v8 and v10 and it's timed out. Thanks!
", "rawMarkdown": "Hello! Can you show your inference code? Because I'm trying to integrate v8 and v10 and it's timed out. Thanks!", "votes": 1, "replies": [ { "id": 3217893, "postDate": "2025-06-05T13:46:27.977Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3217990, "postDate": "2025-06-05T16:23:37.680Z", "content": "We can adjust this parameter CONCENTRATION to accelerate the inference process. Adjusting it to 0.5 will increase the inference speed by 2 times, and adjusting it to 0.25 will increase the inference speed by 4 times.
", "rawMarkdown": "We can adjust this parameter CONCENTRATION to accelerate the inference process. Adjusting it to 0.5 will increase the inference speed by 2 times, and adjusting it to 0.25 will increase the inference speed by 4 times." } ] }, { "id": 3217538, "postDate": "2025-06-05T05:47:22.107Z", "content": "may i know how you ensemble your models?
", "rawMarkdown": "may i know how you ensemble your models?", "votes": 1, "replies": [ { "id": 3217562, "postDate": "2025-06-05T06:19:49.743Z", "content": "I used public code and adjusted the parameters CONFIDENCE_THREHOLD, CONCENTATION, and BATCHSIZE. I also modified a function by changing 'mean' to 'max' in the following code:
\navg_conf = np.mean([d['confidence'] for d in similar_detections])
\nbecause I saw a post in the discussion that 'max' is more stable than 'mean' and 'sum'.
\nhttps://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/578156
Hello! Can you show your inference code? Because I'm trying to integrate v8 and v10 and it's timed out. Thanks!
", "votes": 1, "replies": [ { "id": 3217893, "author_name": "", "author_url": "", "post_date": "2025-06-05T13:46:27.977000", "content": "", "votes": 0, "replies": [] }, { "id": 3217990, "author_name": "", "author_url": "", "post_date": "2025-06-05T16:23:37.680000", "content": "We can adjust this parameter CONCENTRATION to accelerate the inference process. Adjusting it to 0.5 will increase the inference speed by 2 times, and adjusting it to 0.25 will increase the inference speed by 4 times.
", "votes": 0, "replies": [] } ] }, { "id": 3217538, "author_name": "MengYe", "author_url": "", "post_date": "2025-06-05T05:47:22.107000", "content": "may i know how you ensemble your models?
", "votes": 1, "replies": [ { "id": 3217562, "author_name": "", "author_url": "", "post_date": "2025-06-05T06:19:49.743000", "content": "I used public code and adjusted the parameters CONFIDENCE_THREHOLD, CONCENTATION, and BATCHSIZE. I also modified a function by changing 'mean' to 'max' in the following code:
\navg_conf = np.mean([d['confidence'] for d in similar_detections])
\nbecause I saw a post in the discussion that 'max' is more stable than 'mean' and 'sum'.
\nhttps://www.kaggle.com/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/578156