Hi everyone,
\nFor the past month, our team has been only modifying the RANSAC parameters for improving our score. But with this approach, it's been difficult to beat the host's baseline by a good margin. I would love to know what others have tried out so far except changing the RANSAC parameters.
", "messages": [ { "id": "995326", "postDate": "09/02/2020 11:06:33", "content": "Hi everyone,
\nFor the past month, our team has been only modifying the RANSAC parameters for improving our score. But with this approach, it's been difficult to beat the host's baseline by a good margin. I would love to know what others have tried out so far except changing the RANSAC parameters.
", "rawMarkdown": "Hi everyone,\n\nFor the past month, our team has been only modifying the RANSAC parameters for improving our score. But with this approach, it's been difficult to beat the host's baseline by a good margin. I would love to know what others have tried out so far except changing the RANSAC parameters.", "votes": null }, { "id": "995538", "postDate": "09/02/2020 14:33:45", "content": "It is probably the worst way. I dont think you will learn much by just tuning some parameters in a public kernel.
", "rawMarkdown": "It is probably the worst way. I dont think you will learn much by just tuning some parameters in a public kernel.", "votes": null }, { "id": "995600", "postDate": "09/02/2020 15:45:43", "content": "Well-said. There are various sources of value in this competition, RANSAC hyper-parameters being probably the most easily overfitable to the public LB
", "rawMarkdown": "Well-said. There are various sources of value in this competition, RANSAC hyper-parameters being probably the most easily overfitable to the public LB", "votes": null }, { "id": "995627", "postDate": "09/02/2020 16:17:29", "content": "Thanks for your feedback @philippsinger @narsil . Will think about another approach then. Any hints would be appreciated !
", "rawMarkdown": "Thanks for your feedback @philippsinger @narsil . Will think about another approach then. Any hints would be appreciated !", "votes": null }, { "id": "995802", "postDate": "09/02/2020 19:33:17", "content": "What about fitting models?
", "rawMarkdown": "What about fitting models?", "votes": null }, { "id": "995953", "postDate": "09/03/2020 01:41:22", "content": "The data is really really BIG. It's hard to fit for the people without much computing resources.
", "rawMarkdown": "The data is really really BIG. It's hard to fit for the people without much computing resources.", "votes": null }, { "id": "996217", "postDate": "09/03/2020 06:39:29", "content": "The data is not that big. There are so many ways to reduce the fitting time. Be creative and see it as a chance to learn.
", "rawMarkdown": "The data is not that big. There are so many ways to reduce the fitting time. Be creative and see it as a chance to learn.", "votes": null }, { "id": "996353", "postDate": "09/03/2020 08:37:26", "content": "\"worst way\"(?) Can't one do little of both ;) Never used RANSAC hyper-parameters before, is it only overfitting that shows when trying to make the best of the baseline and the parameters? Should one keep the parameters as they were not to ruin the work from the training and testing that came with model?
", "rawMarkdown": "\"worst way\"(?) Can't one do little of both ;) Never used RANSAC hyper-parameters before, is it only overfitting that shows when trying to make the best of the baseline and the parameters? Should one keep the parameters as they were not to ruin the work from the training and testing that came with model?", "votes": null }, { "id": "996597", "postDate": "09/03/2020 12:21:11", "content": "Not really getting your question.
", "rawMarkdown": "Not really getting your question.", "votes": null }, { "id": "996629", "postDate": "09/03/2020 12:46:15", "content": "The first is a comment to the your comment \"worst way\": My comment -Can't one do little of both ;) meaning that one maybe can work with both RANSAC/baseline parameters/tuning and also fitting/training an own model from scratch. One should take wisdom in both was the underlying message, directly or after. \"worst way\" sounded so absolute :) all well, just a comment :)
\nThen my two questions, they are to the general dicussion, with connection to the above, specially the comment \"RANSAC hyper-parameters being probably the most easily overfitable\"
\nTo that I wonder, due to my novice level on the subject/model.
I put together some information about Ransac, from what I have found and read, https://www.kaggle.com/c/landmark-recognition-2020/discussion/180921
", "rawMarkdown": "I put together some information about Ransac, from what I have found and read, https://www.kaggle.com/c/landmark-recognition-2020/discussion/180921", "votes": null }, { "id": "1001223", "postDate": "09/07/2020 06:26:20", "content": "Thanks for sharing!
", "rawMarkdown": "Thanks for sharing!", "votes": null }, { "id": "1002640", "postDate": "09/08/2020 09:54:16", "content": "\n\n\"I have never used RANSAC hyper-parameters before, is it only overfitting that shows when trying to make the best of the baseline and the parameters?\"
\n
You wont know. Thats exactly the problem.
\n\n\n\"Should one keep the parameters as they were not to ruin the work from the training and testing that came with model?\"
\n
No, I would always try to understand and tweak parameters. But you need a saver way to evaluate than the public leaderboard.
", "rawMarkdown": "> \"I have never used RANSAC hyper-parameters before, is it only overfitting that shows when trying to make the best of the baseline and the parameters?\" \n\nYou wont know. Thats exactly the problem. \n\n> \"Should one keep the parameters as they were not to ruin the work from the training and testing that came with model?\"\n\nNo, I would always try to understand and tweak parameters. But you need a saver way to evaluate than the public leaderboard.", "votes": null }, { "id": "1002668", "postDate": "09/08/2020 10:23:02", "content": "Thanks @christofhenkel for the answer. Yes fitting the LB is not a good and reliable way even though previous Landmark competitions have lack of shakeups. Combine with local validation and see correlation is better.
\nJust for my own interest, I have now recently, due to this competition read a lot about Ransac and its area and different implementations, get to take this Landmark Recognition competition which is my first as an experience and education.
\nHave implemented a model that goes with the recommended settings from research papers combined with prev. landmark comp. top teams ransac or similar implementation, doesn't result in the best LB score but if the deadline was now, without any further testing, I would have picked it as one of the two subs.
It is probably the worst way. I dont think you will learn much by just tuning some parameters in a public kernel.
", "votes": null, "replies": [ { "id": 995600, "author_name": "narsil", "author_url": "", "post_date": "09/02/2020 15:45:43", "content": "Well-said. There are various sources of value in this competition, RANSAC hyper-parameters being probably the most easily overfitable to the public LB
", "votes": null, "replies": [] }, { "id": 995627, "author_name": "madz2000", "author_url": "", "post_date": "09/02/2020 16:17:29", "content": "Thanks for your feedback @philippsinger @narsil . Will think about another approach then. Any hints would be appreciated !
", "votes": null, "replies": [] }, { "id": 995802, "author_name": "philippsinger", "author_url": "", "post_date": "09/02/2020 19:33:17", "content": "What about fitting models?
", "votes": null, "replies": [] }, { "id": 995953, "author_name": "gzl0506", "author_url": "", "post_date": "09/03/2020 01:41:22", "content": "The data is really really BIG. It's hard to fit for the people without much computing resources.
", "votes": null, "replies": [] }, { "id": 996217, "author_name": "philippsinger", "author_url": "", "post_date": "09/03/2020 06:39:29", "content": "The data is not that big. There are so many ways to reduce the fitting time. Be creative and see it as a chance to learn.
", "votes": null, "replies": [] }, { "id": 996353, "author_name": "kirderf", "author_url": "", "post_date": "09/03/2020 08:37:26", "content": "\"worst way\"(?) Can't one do little of both ;) Never used RANSAC hyper-parameters before, is it only overfitting that shows when trying to make the best of the baseline and the parameters? Should one keep the parameters as they were not to ruin the work from the training and testing that came with model?
", "votes": null, "replies": [] }, { "id": 996597, "author_name": "philippsinger", "author_url": "", "post_date": "09/03/2020 12:21:11", "content": "Not really getting your question.
", "votes": null, "replies": [] }, { "id": 996629, "author_name": "kirderf", "author_url": "", "post_date": "09/03/2020 12:46:15", "content": "The first is a comment to the your comment \"worst way\": My comment -Can't one do little of both ;) meaning that one maybe can work with both RANSAC/baseline parameters/tuning and also fitting/training an own model from scratch. One should take wisdom in both was the underlying message, directly or after. \"worst way\" sounded so absolute :) all well, just a comment :)
\nThen my two questions, they are to the general dicussion, with connection to the above, specially the comment \"RANSAC hyper-parameters being probably the most easily overfitable\"
\nTo that I wonder, due to my novice level on the subject/model.
\n\n\"I have never used RANSAC hyper-parameters before, is it only overfitting that shows when trying to make the best of the baseline and the parameters?\"
\n
You wont know. Thats exactly the problem.
\n\n\n\"Should one keep the parameters as they were not to ruin the work from the training and testing that came with model?\"
\n
No, I would always try to understand and tweak parameters. But you need a saver way to evaluate than the public leaderboard.
", "votes": null, "replies": [] }, { "id": 1002668, "author_name": "kirderf", "author_url": "", "post_date": "09/08/2020 10:23:02", "content": "Thanks @christofhenkel for the answer. Yes fitting the LB is not a good and reliable way even though previous Landmark competitions have lack of shakeups. Combine with local validation and see correlation is better.
\nJust for my own interest, I have now recently, due to this competition read a lot about Ransac and its area and different implementations, get to take this Landmark Recognition competition which is my first as an experience and education.
\nHave implemented a model that goes with the recommended settings from research papers combined with prev. landmark comp. top teams ransac or similar implementation, doesn't result in the best LB score but if the deadline was now, without any further testing, I would have picked it as one of the two subs.
I put together some information about Ransac, from what I have found and read, https://www.kaggle.com/c/landmark-recognition-2020/discussion/180921
", "votes": null, "replies": [ { "id": 1001223, "author_name": "madz2000", "author_url": "", "post_date": "09/07/2020 06:26:20", "content": "Thanks for sharing!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "995326": "Hi everyone,\n\nFor the past month, our team has been only modifying the RANSAC parameters for improving our score. But with this approach, it's been difficult to beat the host's baseline by a good margin. I would love to know what others have tried out so far except changing the RANSAC parameters.", "995538": "It is probably the worst way. I dont think you will learn much by just tuning some parameters in a public kernel.", "995600": "Well-said. There are various sources of value in this competition, RANSAC hyper-parameters being probably the most easily overfitable to the public LB", "995627": "Thanks for your feedback @philippsinger @narsil . Will think about another approach then. Any hints would be appreciated !", "995802": "What about fitting models?", "995953": "The data is really really BIG. It's hard to fit for the people without much computing resources.", "996217": "The data is not that big. There are so many ways to reduce the fitting time. Be creative and see it as a chance to learn.", "996353": "\"worst way\"(?) Can't one do little of both ;) Never used RANSAC hyper-parameters before, is it only overfitting that shows when trying to make the best of the baseline and the parameters? Should one keep the parameters as they were not to ruin the work from the training and testing that came with model?", "996597": "Not really getting your question.", "996629": "The first is a comment to the your comment \"worst way\": My comment -Can't one do little of both ;) meaning that one maybe can work with both RANSAC/baseline parameters/tuning and also fitting/training an own model from scratch. One should take wisdom in both was the underlying message, directly or after. \"worst way\" sounded so absolute :) all well, just a comment :)\n\nThen my two questions, they are to the general dicussion, with connection to the above, specially the comment \"RANSAC hyper-parameters being probably the most easily overfitable\"\nTo that I wonder, due to my novice level on the subject/model.\n1. \"I have never used RANSAC hyper-parameters before, is it only overfitting that shows when trying to make the best of the baseline and the parameters?\" \n2. \"Should one keep the parameters as they were not to ruin the work from the training and testing that came with model?\"", "1000909": "I put together some information about Ransac, from what I have found and read, https://www.kaggle.com/c/landmark-recognition-2020/discussion/180921", "1001223": "Thanks for sharing!", "1002640": "> \"I have never used RANSAC hyper-parameters before, is it only overfitting that shows when trying to make the best of the baseline and the parameters?\" \n\nYou wont know. Thats exactly the problem. \n\n> \"Should one keep the parameters as they were not to ruin the work from the training and testing that came with model?\"\n\nNo, I would always try to understand and tweak parameters. But you need a saver way to evaluate than the public leaderboard.", "1002668": "Thanks @christofhenkel for the answer. Yes fitting the LB is not a good and reliable way even though previous Landmark competitions have lack of shakeups. Combine with local validation and see correlation is better. \nJust for my own interest, I have now recently, due to this competition read a lot about Ransac and its area and different implementations, get to take this Landmark Recognition competition which is my first as an experience and education.\nHave implemented a model that goes with the recommended settings from research papers combined with prev. landmark comp. top teams ransac or similar implementation, doesn't result in the best LB score but if the deadline was now, without any further testing, I would have picked it as one of the two subs." }, "source": "meta" }