![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2640938%2Fbbb555507832683f3bb53e995bf5380b%2F2.png?generation=1717599897657757&alt=media\")
First, I would like to express my gratitude to the competition organizers and Kaggle staff. I have learned a lot from this competition.
\n
Our basic strategy involves using a rotation-resistant model based on IMC2023 6th place method, along with a multi-model approach using Alike-LightGlue. Moreover, we made several adjustments depending on the scene.
\nOur teammate @sugupoko, discovered that using only the central part of the image improves accuracy for transparent objects. This approach was based on the assumption that transparent objects do not move from the center. As a result, our local score for the cylinder exceeded 0.2. Additionally, by setting the camera model to single-camera, our cylinder score reached 0.463.
\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2640938%2F54a639f705f938781ecde5faa806e1cf%2F3.png?generation=1717600002927540&alt=media\")
We observed that the cylinder overreacted to reflected light and tended to match incorrect keypoints due to minimal image variation. Therefore, after matching, we removed keypoints where the x and y coordinate differences were both within 5 pixels, as these were likely incorrect keypoints.
\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2640938%2F4e690c419a844e454009816936fb7c7f%2F4.png?generation=1717600034499906&alt=media\")
The challenge with church was distinguishing between the front and back, causing the camera to focus on the front.
\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2640938%2F30cc61a03702e1107976131439cd5859%2F5.png?generation=1717600111598243&alt=media\")
\nTo capture finer details beyond just the clock, we implemented an approach to detect keypoints by dividing the image into four sections. Ultimately, with a model ensemble and setting the camera model to simple-pinhole, we achieved a score of 0.3561 on the train data.
\n(Note: Our final submission used simple-radial for all scenes.)
\nHowever, this alone did not resolve the issue of distinguishing between the front and back as mentioned above.
First, I would like to express my gratitude to the competition organizers and Kaggle staff. I have learned a lot from this competition.
\n
Our basic strategy involves using a rotation-resistant model based on IMC2023 6th place method, along with a multi-model approach using Alike-LightGlue. Moreover, we made several adjustments depending on the scene.
\nOur teammate @sugupoko, discovered that using only the central part of the image improves accuracy for transparent objects. This approach was based on the assumption that transparent objects do not move from the center. As a result, our local score for the cylinder exceeded 0.2. Additionally, by setting the camera model to single-camera, our cylinder score reached 0.463.
\n
We observed that the cylinder overreacted to reflected light and tended to match incorrect keypoints due to minimal image variation. Therefore, after matching, we removed keypoints where the x and y coordinate differences were both within 5 pixels, as these were likely incorrect keypoints.
\n
The challenge with church was distinguishing between the front and back, causing the camera to focus on the front.
\n
\nTo capture finer details beyond just the clock, we implemented an approach to detect keypoints by dividing the image into four sections. Ultimately, with a model ensemble and setting the camera model to simple-pinhole, we achieved a score of 0.3561 on the train data.
\n(Note: Our final submission used simple-radial for all scenes.)
\nHowever, this alone did not resolve the issue of distinguishing between the front and back as mentioned above.