![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4495635%2F938578d75aaafd80ef0108855535140a%2FIMC.jpg?generation=1713187931007916&alt=media\")
Hello, Kagglers!
\nI'm sharing the IMC2024 Starter, which gives a consistent leaderboard score of 0.13. This starter provides a solid base to build upon.
\nWe can explore fine-tuning hyperparameters, modifying the code for additional image processing techniques, and trying different models for feature extraction. These avenues offer opportunities to potentially improve the performance further.
\nLet me know if there are issues or if you have any questions.!
\n
Hello, Kagglers!
\nI'm sharing the IMC2024 Starter, which gives a consistent leaderboard score of 0.13. This starter provides a solid base to build upon.
\nWe can explore fine-tuning hyperparameters, modifying the code for additional image processing techniques, and trying different models for feature extraction. These avenues offer opportunities to potentially improve the performance further.
\nLet me know if there are issues or if you have any questions.!
\n
Thanks for sharing! I will have to check this out! This competition was hard last year, this time around it is even more difficult!
", "rawMarkdown": "Thanks for sharing! I will have to check this out! This competition was hard last year, this time around it is even more difficult!", "votes": null }, { "id": "2753496", "postDate": "04/15/2024 14:22:09", "content": "You are very welcome! it's a tricky one for sure.
", "rawMarkdown": "You are very welcome! it's a tricky one for sure.", "votes": null }, { "id": "2756672", "postDate": "04/17/2024 06:08:25", "content": "Thank you very much for your sharing!
\nI have a question: is the \"cross-validation score\" that you mentioned derived from this metric? Because the validation metrics I obtained differ greatly from the leaderboard score.
You are very welcome!
\n\n\nI have a question: is the \"cross-validation score\" that you mentioned derived from this metric?
\n
Yes, it's the same.
\nRandomness is an issue for scoring, but can be fixed by setting thread_num=1 for pycolmap, and you could calibrate the number of images in a dataset for cross validation until you get a similar LB score.
\nCould you share the numbers for CV and LB?
", "rawMarkdown": "You are very welcome!\n\n> I have a question: is the \"cross-validation score\" that you mentioned derived from this metric?\n\nYes, it's the same.\n\nRandomness is an issue for scoring, but can be fixed by setting thread_num=1 for pycolmap, and you could calibrate the number of images in a dataset for cross validation until you get a similar LB score.\n\nCould you share the numbers for CV and LB?", "votes": null }, { "id": "2757557", "postDate": "04/17/2024 15:21:44", "content": "Thank you for your suggestion!
\nIn early submissions, I got a CV score of 0.39, but only 0.13 on the LB
Two related errors in first block. By any chance can you provide the requirements.txt file?
\nERROR: Could not open requirements file: [Errno 2] No such file or directory: '/kaggle/input/check-image-orientation/requirements.txt'
\nModuleNotFoundError Traceback (most recent call last)
\nCell In[3], line 1
\n----> 1 from imc24 import *
\n 2 sys.path.append(\"/kaggle/input/colmap-db-import\")
\n 3 from database import *
File /kaggle/usr/lib/imc24/imc24.py:40
\n 38 from lightglue import LightGlue, ALIKED, match_pair
\n 39 from transformers import AutoImageProcessor, AutoModel
\n---> 40 from check_orientation.pre_trained_models import create_model
\n 42 IMC_PATH = '/kaggle/input/image-matching-challenge-2024'
\n 43 DEVICE = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")
ModuleNotFoundError: No module named 'check_orientation
", "rawMarkdown": "Two related errors in first block. By any chance can you provide the requirements.txt file?\n\nERROR: Could not open requirements file: [Errno 2] No such file or directory: '/kaggle/input/check-image-orientation/requirements.txt'\n\n---------------------------------------------------------------------------\nModuleNotFoundError Traceback (most recent call last)\nCell In[3], line 1\n----> 1 from imc24 import *\n 2 sys.path.append(\"/kaggle/input/colmap-db-import\")\n 3 from database import *\n\nFile /kaggle/usr/lib/imc24/imc24.py:40\n 38 from lightglue import LightGlue, ALIKED, match_pair\n 39 from transformers import AutoImageProcessor, AutoModel\n---> 40 from check_orientation.pre_trained_models import create_model\n 42 IMC_PATH = '/kaggle/input/image-matching-challenge-2024'\n 43 DEVICE = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n\nModuleNotFoundError: No module named 'check_orientation", "votes": null }, { "id": "2758793", "postDate": "04/18/2024 10:56:09", "content": "Sure thing, I just updated the code.
\nTwo things to keep in mind:
So you would have to modify the rotation function to include specific datasets, the dioscuri dataset seems to work well with it.
", "rawMarkdown": "Sure thing, I just updated the code. \nTwo things to keep in mind:\n- The image rotation does not work for transparent objects(cup and cylinder).\n- If you decide to rotate all the datasets, the official scoring runs out of time.\n\nSo you would have to modify the rotation function to include specific datasets, the dioscuri dataset seems to work well with it.", "votes": null } ], "comments": [ { "id": 2753485, "author_name": "cody11null", "author_url": "", "post_date": "04/15/2024 14:06:47", "content": "Thanks for sharing! I will have to check this out! This competition was hard last year, this time around it is even more difficult!
", "votes": null, "replies": [ { "id": 2753496, "author_name": "nartaa", "author_url": "", "post_date": "04/15/2024 14:22:09", "content": "You are very welcome! it's a tricky one for sure.
", "votes": null, "replies": [] } ] }, { "id": 2756672, "author_name": "sunnyykk", "author_url": "", "post_date": "04/17/2024 06:08:25", "content": "Thank you very much for your sharing!
\nI have a question: is the \"cross-validation score\" that you mentioned derived from this metric? Because the validation metrics I obtained differ greatly from the leaderboard score.
You are very welcome!
\n\n\nI have a question: is the \"cross-validation score\" that you mentioned derived from this metric?
\n
Yes, it's the same.
\nRandomness is an issue for scoring, but can be fixed by setting thread_num=1 for pycolmap, and you could calibrate the number of images in a dataset for cross validation until you get a similar LB score.
\nCould you share the numbers for CV and LB?
", "votes": null, "replies": [ { "id": 2757557, "author_name": "sunnyykk", "author_url": "", "post_date": "04/17/2024 15:21:44", "content": "Thank you for your suggestion!
\nIn early submissions, I got a CV score of 0.39, but only 0.13 on the LB
Two related errors in first block. By any chance can you provide the requirements.txt file?
\nERROR: Could not open requirements file: [Errno 2] No such file or directory: '/kaggle/input/check-image-orientation/requirements.txt'
\nModuleNotFoundError Traceback (most recent call last)
\nCell In[3], line 1
\n----> 1 from imc24 import *
\n 2 sys.path.append(\"/kaggle/input/colmap-db-import\")
\n 3 from database import *
File /kaggle/usr/lib/imc24/imc24.py:40
\n 38 from lightglue import LightGlue, ALIKED, match_pair
\n 39 from transformers import AutoImageProcessor, AutoModel
\n---> 40 from check_orientation.pre_trained_models import create_model
\n 42 IMC_PATH = '/kaggle/input/image-matching-challenge-2024'
\n 43 DEVICE = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")
ModuleNotFoundError: No module named 'check_orientation
", "votes": null, "replies": [ { "id": 2758793, "author_name": "nartaa", "author_url": "", "post_date": "04/18/2024 10:56:09", "content": "Sure thing, I just updated the code.
\nTwo things to keep in mind:
So you would have to modify the rotation function to include specific datasets, the dioscuri dataset seems to work well with it.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2753418": "Hello, Kagglers!\n\nI'm sharing the [IMC2024 Starter](https://www.kaggle.com/code/nartaa/imc2024-starter), which gives a consistent leaderboard score of 0.13. This starter provides a solid base to build upon.\n\nWe can explore fine-tuning hyperparameters, modifying the code for additional image processing techniques, and trying different models for feature extraction. These avenues offer opportunities to potentially improve the performance further.\n\nLet me know if there are issues or if you have any questions.!\n\n", "2753485": "Thanks for sharing! I will have to check this out! This competition was hard last year, this time around it is even more difficult!", "2753496": "You are very welcome! it's a tricky one for sure.", "2756672": "Thank you very much for your sharing! \nI have a question: is the \"cross-validation score\" that you mentioned derived from [this metric](https://www.kaggle.com/code/fabiobellavia/imc2024-3d-metric-evaluation-example)? Because the validation metrics I obtained differ greatly from the leaderboard score.", "2756825": "You are very welcome!\n\n> I have a question: is the \"cross-validation score\" that you mentioned derived from this metric?\n\nYes, it's the same.\n\nRandomness is an issue for scoring, but can be fixed by setting thread_num=1 for pycolmap, and you could calibrate the number of images in a dataset for cross validation until you get a similar LB score.\n\nCould you share the numbers for CV and LB?", "2757557": "Thank you for your suggestion! \nIn early submissions, I got a CV score of 0.39, but only 0.13 on the LB", "2758542": "Two related errors in first block. By any chance can you provide the requirements.txt file?\n\nERROR: Could not open requirements file: [Errno 2] No such file or directory: '/kaggle/input/check-image-orientation/requirements.txt'\n\n---------------------------------------------------------------------------\nModuleNotFoundError Traceback (most recent call last)\nCell In[3], line 1\n----> 1 from imc24 import *\n 2 sys.path.append(\"/kaggle/input/colmap-db-import\")\n 3 from database import *\n\nFile /kaggle/usr/lib/imc24/imc24.py:40\n 38 from lightglue import LightGlue, ALIKED, match_pair\n 39 from transformers import AutoImageProcessor, AutoModel\n---> 40 from check_orientation.pre_trained_models import create_model\n 42 IMC_PATH = '/kaggle/input/image-matching-challenge-2024'\n 43 DEVICE = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n\nModuleNotFoundError: No module named 'check_orientation", "2758793": "Sure thing, I just updated the code. \nTwo things to keep in mind:\n- The image rotation does not work for transparent objects(cup and cylinder).\n- If you decide to rotate all the datasets, the official scoring runs out of time.\n\nSo you would have to modify the rotation function to include specific datasets, the dioscuri dataset seems to work well with it." }, "source": "meta" }