{ "id": 501721, "title": "Shiny & glass objects", "url": "/competitions/image-matching-challenge-2024/discussion/501721", "author_name": "Bhavesh Jain", "post_date": "2024-05-10T13:18:08.471000", "votes": 18, "comment_count": 10, "views": 0, "content": "

i am genuinely confused what can be done about this category of images
\ni plotted them & there is nothing differentiable for even human eyes
\nhow and what even are we going to use to do 3d reconstruction from these intelligible images
\nAnyone care to jump in?

", "messages": [ { "id": 2805278, "postDate": "2024-05-10T13:18:08.470Z", "content": "

i am genuinely confused what can be done about this category of images
\ni plotted them & there is nothing differentiable for even human eyes
\nhow and what even are we going to use to do 3d reconstruction from these intelligible images
\nAnyone care to jump in?

", "rawMarkdown": "i am genuinely confused what can be done about this category of images\ni plotted them & there is nothing differentiable for even human eyes\nhow and what even are we going to use to do 3d reconstruction from these intelligible images\nAnyone care to jump in?", "votes": 17 }, { "id": 2805299, "postDate": "2024-05-10T13:43:29.920Z", "content": "

I feel like I'm in the same boat as you, the images themselves are very confusing. Plus the nature based images feel like they want a 360 degree reconstruction rather than a basic 3D model.

", "rawMarkdown": "I feel like I'm in the same boat as you, the images themselves are very confusing. Plus the nature based images feel like they want a 360 degree reconstruction rather than a basic 3D model.", "votes": 3, "replies": [ { "id": 2805308, "postDate": "2024-05-10T13:49:30.497Z", "content": "

I thought I was the only one who is lost😂
\nYes, nature images seem to be even more of a lost cause

", "rawMarkdown": "I thought I was the only one who is lost😂\nYes, nature images seem to be even more of a lost cause", "votes": 3 } ] }, { "id": 2806047, "postDate": "2024-05-10T20:41:09.700Z", "content": "

I have same thoughts, not sure what to do with those, and the background for the cylinder images is blurry, so no information to get from there.

", "rawMarkdown": "I have same thoughts, not sure what to do with those, and the background for the cylinder images is blurry, so no information to get from there.", "votes": 1, "replies": [ { "id": 2806070, "postDate": "2024-05-10T21:09:38.880Z", "content": "

@nartaa I had some improvement using CLAHE with a large cliplimit value when using the training dataset and applying it to only the 'transparencies' category….unfortunately the result seems to be the opposite on the LB….so no luck with that yet.

\n

Still a lot of keypoints are generated around the really shiny and reflected areas. I think if that could be trimmed down than a lot more accurate keypoints remain.

\n

But for now that's just theory ;-)

", "rawMarkdown": "@nartaa I had some improvement using CLAHE with a large cliplimit value when using the training dataset and applying it to only the 'transparencies' category....unfortunately the result seems to be the opposite on the LB....so no luck with that yet.\n\nStill a lot of keypoints are generated around the really shiny and reflected areas. I think if that could be trimmed down than a lot more accurate keypoints remain.\n\nBut for now that's just theory ;-)", "votes": 3, "replies": [ { "id": 2806098, "postDate": "2024-05-10T21:43:27.833Z", "content": "

I saw all the images, and I have no clue, what are the difference between any of the images & so I find it hard to believe that any keypoints (if found) would be meaningful

", "rawMarkdown": "I saw all the images, and I have no clue, what are the difference between any of the images & so I find it hard to believe that any keypoints (if found) would be meaningful ", "votes": 2 }, { "id": 2806121, "postDate": "2024-05-10T22:06:35.540Z", "content": "

@bhavesjain On the glass cylinder you can see fingerprints, scratches and some rotating, mounting structure on the left.

\n

On the right there are however some white shiny areas that look like light reflections…they appear to remain mostly in place.

\n

I 'think' that is causing a lot of noisy keypoints that limit the results that can be achieved.

\n

But on how to improve that … no clue. CLAHE sometimes worked for me on the training set…not on the LB testset. So I'am also stuck there….just sharing my results with it sofar :-)

", "rawMarkdown": "@bhavesjain On the glass cylinder you can see fingerprints, scratches and some rotating, mounting structure on the left.\n\nOn the right there are however some white shiny areas that look like light reflections...they appear to remain mostly in place.\n\nI 'think' that is causing a lot of noisy keypoints that limit the results that can be achieved.\n\nBut on how to improve that ... no clue. CLAHE sometimes worked for me on the training set...not on the LB testset. So I'am also stuck there....just sharing my results with it sofar :-)\n", "votes": 2 }, { "id": 2806122, "postDate": "2024-05-10T22:12:15.013Z", "content": "

I too was thinking that some minor scratches(if present) could be used, but until you pointed it out, I wasn't sure if they'd be enough to do a 3d reconstruction

", "rawMarkdown": "I too was thinking that some minor scratches(if present) could be used, but until you pointed it out, I wasn't sure if they'd be enough to do a 3d reconstruction ", "votes": 1 }, { "id": 2806281, "postDate": "2024-05-11T02:25:09.717Z", "content": "

@rsmits Thanks for sharing.

", "rawMarkdown": "@rsmits Thanks for sharing.", "votes": 1 }, { "id": 2823485, "postDate": "2024-05-19T08:25:07.660Z", "content": "

I am confused -- what should be the expected camera poses of the transparent category? The camera is obviously static with respect to the environment, which means all groundtruth camera poses should be the same -- this is aligned with the groundtruth camera poses provided. However, does it mean that if we encounter any of such cases during the test phase, we can just predict the camera to be static and can get the full score this way?

", "rawMarkdown": "I am confused -- what should be the expected camera poses of the transparent category? The camera is obviously static with respect to the environment, which means all groundtruth camera poses should be the same -- this is aligned with the groundtruth camera poses provided. However, does it mean that if we encounter any of such cases during the test phase, we can just predict the camera to be static and can get the full score this way? ", "votes": 1 }, { "id": 2827262, "postDate": "2024-05-21T10:58:14.900Z", "rawMarkdown": "", "isDeleted": true } ] } ] } ], "comments": [ { "id": 2805299, "author_name": "Prithvi Gulati", "author_url": "", "post_date": "2024-05-10T13:43:29.920000", "content": "

I feel like I'm in the same boat as you, the images themselves are very confusing. Plus the nature based images feel like they want a 360 degree reconstruction rather than a basic 3D model.

", "votes": 3, "replies": [ { "id": 2805308, "author_name": "Bhavesh Jain", "author_url": "", "post_date": "2024-05-10T13:49:30.497000", "content": "

I thought I was the only one who is lost😂
\nYes, nature images seem to be even more of a lost cause

", "votes": 3, "replies": [] } ] }, { "id": 2806047, "author_name": "Danial Zakaria", "author_url": "", "post_date": "2024-05-10T20:41:09.700000", "content": "

I have same thoughts, not sure what to do with those, and the background for the cylinder images is blurry, so no information to get from there.

", "votes": 1, "replies": [ { "id": 2806070, "author_name": "Robin Smits", "author_url": "", "post_date": "2024-05-10T21:09:38.880000", "content": "

@nartaa I had some improvement using CLAHE with a large cliplimit value when using the training dataset and applying it to only the 'transparencies' category….unfortunately the result seems to be the opposite on the LB….so no luck with that yet.

\n

Still a lot of keypoints are generated around the really shiny and reflected areas. I think if that could be trimmed down than a lot more accurate keypoints remain.

\n

But for now that's just theory ;-)

", "votes": 3, "replies": [ { "id": 2806098, "author_name": "Bhavesh Jain", "author_url": "", "post_date": "2024-05-10T21:43:27.833000", "content": "

I saw all the images, and I have no clue, what are the difference between any of the images & so I find it hard to believe that any keypoints (if found) would be meaningful

", "votes": 2, "replies": [] }, { "id": 2806121, "author_name": "Robin Smits", "author_url": "", "post_date": "2024-05-10T22:06:35.540000", "content": "

@bhavesjain On the glass cylinder you can see fingerprints, scratches and some rotating, mounting structure on the left.

\n

On the right there are however some white shiny areas that look like light reflections…they appear to remain mostly in place.

\n

I 'think' that is causing a lot of noisy keypoints that limit the results that can be achieved.

\n

But on how to improve that … no clue. CLAHE sometimes worked for me on the training set…not on the LB testset. So I'am also stuck there….just sharing my results with it sofar :-)

", "votes": 2, "replies": [] }, { "id": 2806122, "author_name": "Bhavesh Jain", "author_url": "", "post_date": "2024-05-10T22:12:15.013000", "content": "

I too was thinking that some minor scratches(if present) could be used, but until you pointed it out, I wasn't sure if they'd be enough to do a 3d reconstruction

", "votes": 1, "replies": [] }, { "id": 2806281, "author_name": "Danial Zakaria", "author_url": "", "post_date": "2024-05-11T02:25:09.717000", "content": "

@rsmits Thanks for sharing.

", "votes": 1, "replies": [] }, { "id": 2823485, "author_name": "Yuxiang Huang", "author_url": "", "post_date": "2024-05-19T08:25:07.660000", "content": "

I am confused -- what should be the expected camera poses of the transparent category? The camera is obviously static with respect to the environment, which means all groundtruth camera poses should be the same -- this is aligned with the groundtruth camera poses provided. However, does it mean that if we encounter any of such cases during the test phase, we can just predict the camera to be static and can get the full score this way?

", "votes": 1, "replies": [] }, { "id": 2827262, "author_name": "", "author_url": "", "post_date": "2024-05-21T10:58:14.900000", "content": "", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2805278": "i am genuinely confused what can be done about this category of images\ni plotted them & there is nothing differentiable for even human eyes\nhow and what even are we going to use to do 3d reconstruction from these intelligible images\nAnyone care to jump in?", "2805299": "I feel like I'm in the same boat as you, the images themselves are very confusing. Plus the nature based images feel like they want a 360 degree reconstruction rather than a basic 3D model.", "2806047": "I have same thoughts, not sure what to do with those, and the background for the cylinder images is blurry, so no information to get from there." } }