š Hello everyone, in this writeup our team tell about our solution.
\nPreprocessing: Preprocessing notebook. Firstly we merge all datasets that we generated and then delete outlies(Some synthetic photos were inside objects or had other problems).
result solution: ensemble YOLO 8-12. In our result solution we learn yolo8, 11, 12 and merge results at this models at different sizes: image_sizes = [960, 1024, 1440, 2272].
This is actually a very interesting solution, since the dimensions are smaller than the main test image and when we reduce the size, we still lose useful information, but this solution better than sliding window because there are classes in the entire image(candle class at full image) and the sliding window will often make mistakes. I called this method R-CNN because we are looking for small objects (that is, we suspect that these pixels may be, for example, a candle) and then an ensemble of different YOLO aligns the result and, using weighted_boxes_fusion, votes for the most likely results.
\nā bonus: Private score 1.
", "messages": [ { "id": "3384016", "postDate": "12/31/2025 10:03:40", "content": "š Hello everyone, in this writeup our team tell about our solution.
\nPreprocessing: Preprocessing notebook. Firstly we merge all datasets that we generated and then delete outlies(Some synthetic photos were inside objects or had other problems).
result solution: ensemble YOLO 8-12. In our result solution we learn yolo8, 11, 12 and merge results at this models at different sizes: image_sizes = [960, 1024, 1440, 2272].
This is actually a very interesting solution, since the dimensions are smaller than the main test image and when we reduce the size, we still lose useful information, but this solution better than sliding window because there are classes in the entire image(candle class at full image) and the sliding window will often make mistakes. I called this method R-CNN because we are looking for small objects (that is, we suspect that these pixels may be, for example, a candle) and then an ensemble of different YOLO aligns the result and, using weighted_boxes_fusion, votes for the most likely results.
\nā bonus: Private score 1.
", "rawMarkdown": "š Hello everyone, in this writeup our team tell about our solution.\n\n1. Preprocessing: [Preprocessing notebook](https://www.kaggle.com/code/antonoof/data-to-delete-outliers/notebook?scriptVersionId=260727584). Firstly we merge all datasets that we generated and then delete outlies(Some synthetic photos were inside objects or had other problems).\n\n2. result solution: [ensemble YOLO 8-12](https://www.kaggle.com/code/tmdofi/predict-ensemble-yolo/notebook?scriptVersionId=269263680). In our result solution we learn yolo8, 11, 12 and merge results at this models at different sizes: image_sizes = [960, 1024, 1440, 2272]. \n\nThis is actually a very interesting solution, since the dimensions are smaller than the main test image and when we reduce the size, we still lose useful information, but this solution better than sliding window because there are classes in the entire image(candle class at full image) and the sliding window will often make mistakes. I called this method R-CNN because we are looking for small objects (that is, we suspect that these pixels may be, for example, a candle) and then an ensemble of different YOLO aligns the result and, using weighted_boxes_fusion, votes for the most likely results.\n\nā bonus: [Private score 1.](https://www.kaggle.com/code/antonoof/predict-ensemble/notebook?scriptVersionId=264152191)", "votes": null }, { "id": "3386780", "postDate": "01/05/2026 20:23:20", "content": "Fantastic strategy, thank you for sharing!
", "rawMarkdown": "Fantastic strategy, thank you for sharing!", "votes": null } ], "comments": [ { "id": 3386780, "author_name": "rebekahduality", "author_url": "", "post_date": "01/05/2026 20:23:20", "content": "Fantastic strategy, thank you for sharing!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3384016": "š Hello everyone, in this writeup our team tell about our solution.\n\n1. Preprocessing: [Preprocessing notebook](https://www.kaggle.com/code/antonoof/data-to-delete-outliers/notebook?scriptVersionId=260727584). Firstly we merge all datasets that we generated and then delete outlies(Some synthetic photos were inside objects or had other problems).\n\n2. result solution: [ensemble YOLO 8-12](https://www.kaggle.com/code/tmdofi/predict-ensemble-yolo/notebook?scriptVersionId=269263680). In our result solution we learn yolo8, 11, 12 and merge results at this models at different sizes: image_sizes = [960, 1024, 1440, 2272]. \n\nThis is actually a very interesting solution, since the dimensions are smaller than the main test image and when we reduce the size, we still lose useful information, but this solution better than sliding window because there are classes in the entire image(candle class at full image) and the sliding window will often make mistakes. I called this method R-CNN because we are looking for small objects (that is, we suspect that these pixels may be, for example, a candle) and then an ensemble of different YOLO aligns the result and, using weighted_boxes_fusion, votes for the most likely results.\n\nā bonus: [Private score 1.](https://www.kaggle.com/code/antonoof/predict-ensemble/notebook?scriptVersionId=264152191)", "3386780": "Fantastic strategy, thank you for sharing!" }, "source": "meta" }