Hi everyone,
\nWelcome you to the 2025 Image Matching Challenge! We're back, for the fourth time!
\nIn the first edition you asked you to match two images taken from different viewpoints, to figure out the position of one camera with respect to the other. The year after we expanded the competition to multiview 3D reconstruction, with hundreds of images per scene. Last year we raised the bar, asking you to solve 3D reconstruction over six different domains, including urban and natural scenarios and images captured from drones.
\nThis year you'll face a different challenge in 3D reconstruction: we won't tell you which images belong together! This is a problem we encounter very frequently in the real world, where images are grouped together using noisy signals like GPS or text tags.
\nWe provide a few resources to get you started:
\nYou can find many more resources below:
\nNews on prize-eligibility: in previous editions, competitors were not able to use submissions using non-commercial licensed 3rd party code (such as SuperPoint of SuperGlue) and claim prizes. This is now possible, as long as you open-source your code and meet all the conditions (see the rules tab for details).
\nOn behalf of all the organizers, welcome again, and best of luck!
", "messages": [ { "id": "3163251", "postDate": "03/30/2025 13:46:44", "content": "Hi everyone,
\nWelcome you to the 2025 Image Matching Challenge! We're back, for the fourth time!
\nIn the first edition you asked you to match two images taken from different viewpoints, to figure out the position of one camera with respect to the other. The year after we expanded the competition to multiview 3D reconstruction, with hundreds of images per scene. Last year we raised the bar, asking you to solve 3D reconstruction over six different domains, including urban and natural scenarios and images captured from drones.
\nThis year you'll face a different challenge in 3D reconstruction: we won't tell you which images belong together! This is a problem we encounter very frequently in the real world, where images are grouped together using noisy signals like GPS or text tags.
\nWe provide a few resources to get you started:
\nYou can find many more resources below:
\nNews on prize-eligibility: in previous editions, competitors were not able to use submissions using non-commercial licensed 3rd party code (such as SuperPoint of SuperGlue) and claim prizes. This is now possible, as long as you open-source your code and meet all the conditions (see the rules tab for details).
\nOn behalf of all the organizers, welcome again, and best of luck!
", "rawMarkdown": "Hi everyone,\n\nWelcome you to the 2025 Image Matching Challenge! We're back, for the fourth time!\n\nIn the first edition you asked you to match two images taken from different viewpoints, to figure out the position of one camera with respect to the other. The year after we expanded the competition to multiview 3D reconstruction, with hundreds of images per scene. Last year we raised the bar, asking you to solve 3D reconstruction over six different domains, including urban and natural scenarios and images captured from drones.\n\nThis year you'll face a different challenge in 3D reconstruction: we won't tell you which images belong together! This is a problem we encounter very frequently in the real world, where images are grouped together using noisy signals like GPS or text tags.\n\nWe provide a few resources to get you started:\n* A [notebook](https://www.kaggle.com/code/eduardtrulls/imc25-submission/) to create a simple submission on GPU, using ALIKED local features and LightGlue for matching.\n* A [library](https://www.kaggle.com/datasets/eduardtrulls/imc25-utils) with utility functions to interact with colmap and to compute the metric.\n\nYou can find many more resources below:\n* The [CVPR workshop](https://image-matching-workshop.github.io/) we organize along with this challenge.\n* Previous editions of the competition: [2022](https://www.kaggle.com/competitions/image-matching-challenge-2022), [2023](https://www.kaggle.com/competitions/image-matching-challenge-2023), [2024](https://www.kaggle.com/competitions/image-matching-challenge-2024).\n* Our [IJCV paper](https://arxiv.org/abs/2003.01587) on image matching and 3D reconstruction.\n\nNews on prize-eligibility: in previous editions, competitors were not able to use submissions using non-commercial licensed 3rd party code (such as SuperPoint of SuperGlue) and claim prizes. This is now possible, as long as you open-source your code and meet all the conditions (see the [rules](https://www.kaggle.com/competitions/image-matching-challenge-2025/rules) tab for details).\n\nOn behalf of all the organizers, welcome again, and best of luck!", "votes": null }, { "id": "3169435", "postDate": "04/03/2025 13:18:56", "content": "Hello , i'm Ali, Can I use TUP in Kaggle competitions, or is it not allowed ?
", "rawMarkdown": "Hello , i'm Ali, Can I use TUP in Kaggle competitions, or is it not allowed ?", "votes": null }, { "id": "3180088", "postDate": "04/16/2025 06:09:48", "content": "Could we use pretrained model and weights.th files?
", "rawMarkdown": "Could we use pretrained model and weights.th files?", "votes": null }, { "id": "3187596", "postDate": "04/26/2025 09:00:59", "content": "Can we use pretrained models?
", "rawMarkdown": "Can we use pretrained models?", "votes": null }, { "id": "3190309", "postDate": "04/30/2025 13:17:51", "content": "Is it allowed to use pre-trained models?
", "rawMarkdown": "Is it allowed to use pre-trained models?", "votes": null }, { "id": "3193285", "postDate": "05/04/2025 07:09:25", "content": "great work!!
", "rawMarkdown": "great work!!", "votes": null }, { "id": "3201507", "postDate": "05/14/2025 01:14:13", "content": "hi i didnt quite understood what the input amd output should ,anyone please help…
", "rawMarkdown": "hi i didnt quite understood what the input amd output should ,anyone please help...", "votes": null }, { "id": "3202606", "postDate": "05/15/2025 16:24:46", "content": "i just registered so still trying to figure things out,
\ninputs are images from datasets (for example fbk_vineyard) and you need to match them if they belongs to the scene or not and return camera's orientation and translation for each image
Hello , i'm Ali, Can I use TUP in Kaggle competitions, or is it not allowed ?
", "votes": null, "replies": [] }, { "id": 3180088, "author_name": "kevinconnor", "author_url": "", "post_date": "04/16/2025 06:09:48", "content": "Could we use pretrained model and weights.th files?
", "votes": null, "replies": [] }, { "id": 3187596, "author_name": "mohanapavanbezawada", "author_url": "", "post_date": "04/26/2025 09:00:59", "content": "Can we use pretrained models?
", "votes": null, "replies": [] }, { "id": 3190309, "author_name": "rashiedmatan", "author_url": "", "post_date": "04/30/2025 13:17:51", "content": "Is it allowed to use pre-trained models?
", "votes": null, "replies": [] }, { "id": 3193285, "author_name": "ashreenk3", "author_url": "", "post_date": "05/04/2025 07:09:25", "content": "great work!!
", "votes": null, "replies": [] }, { "id": 3201507, "author_name": "ashiraman", "author_url": "", "post_date": "05/14/2025 01:14:13", "content": "hi i didnt quite understood what the input amd output should ,anyone please help…
", "votes": null, "replies": [ { "id": 3202606, "author_name": "nirmitsolanki", "author_url": "", "post_date": "05/15/2025 16:24:46", "content": "i just registered so still trying to figure things out,
\ninputs are images from datasets (for example fbk_vineyard) and you need to match them if they belongs to the scene or not and return camera's orientation and translation for each image