\n
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F18684623%2Fe19f26576512dc7158ed5fbaf25d5dbb%2F1.png?generation=1717738328784104&alt=media\")
Thanks to the organizers for the valuable contest opportunity, and the very good baseline code that gave an avenue for a quick introduction to image matching.
\nMy core program is based on last year's first, second and fifth winning programs. I use aliked, affnet to extract keypoints and perform keypoint fusion. I tried last year's rotate,tweak AdaLAM etc based on training set. I feel the result is very little, also may be some randomness effect.
Thanks to the organizers for the valuable contest opportunity, and the very good baseline code that gave an avenue for a quick introduction to image matching.
\nMy core program is based on last year's first, second and fifth winning programs. I use aliked, affnet to extract keypoints and perform keypoint fusion. I tried last year's rotate,tweak AdaLAM etc based on training set. I feel the result is very little, also may be some randomness effect.
Thanks for sharing.
\nDoes this mean that you changed only ensemble part to get high score ?
\nIs there any pre / post processing ?
In ensemble, I think the combination is so important.
\nHow do you find the combination of models ?
We referred to the previous first, second, and fifth place solutions and attempted to replicate them, but were unable to successfully replicate the first two place solutions. Therefore, we chose the fifth place solution and adjusted the parameters of aligned and affnet.
", "rawMarkdown": "We referred to the previous first, second, and fifth place solutions and attempted to replicate them, but were unable to successfully replicate the first two place solutions. Therefore, we chose the fifth place solution and adjusted the parameters of aligned and affnet.", "votes": null } ], "comments": [ { "id": 2859638, "author_name": "clearwaterkzk", "author_url": "", "post_date": "06/07/2024 06:39:36", "content": "Thanks for sharing.
\nDoes this mean that you changed only ensemble part to get high score ?
\nIs there any pre / post processing ?
In ensemble, I think the combination is so important.
\nHow do you find the combination of models ?
We referred to the previous first, second, and fifth place solutions and attempted to replicate them, but were unable to successfully replicate the first two place solutions. Therefore, we chose the fifth place solution and adjusted the parameters of aligned and affnet.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2859573": "Thanks to the organizers for the valuable contest opportunity, and the very good baseline code that gave an avenue for a quick introduction to image matching.\n \nMy core program is based on [last year's first](https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/417407), [second](https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/416873) and [fifth](https://www.kaggle.com/competitions/image-matching-challenge-2023/discussion/417045) winning programs. I use aliked, affnet to extract keypoints and perform keypoint fusion. I tried last year's rotate,tweak AdaLAM etc based on training set. I feel the result is very little, also may be some randomness effect.