{ "id": 573197, "title": "Whats the highest Yolo LB score?", "url": "/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/573197", "author_name": "m4nocha", "post_date": "2025-04-14T07:11:00.341000", "votes": 11, "comment_count": 20, "views": 0, "content": "

I was wondering whats the highest public LB score has YOLO achieved till now?

", "messages": [ { "id": 3179012, "postDate": "2025-04-14T22:17:05.267Z", "content": "

My LB0.800 is a single model of Yolo

\n\n

There is some ingenuity in post processing
\nBut I didn't use TTA.

", "rawMarkdown": "My LB0.800 is a single model of Yolo \n- Bigger models do not necessarily score better \n- Newer versions do not necessarily score better \nI know that\n\nThere is some ingenuity in post processing\nBut I didn't use TTA.", "votes": 17, "replies": [ { "id": 3179032, "postDate": "2025-04-14T23:40:02.943Z", "content": "

Impressive solo score, post processing is also something I found to require a lot of attention and creativity

", "rawMarkdown": "Impressive solo score, post processing is also something I found to require a lot of attention and creativity", "votes": 1 }, { "id": 3179112, "postDate": "2025-04-15T03:56:31.227Z", "content": "

May I ask if 0.800 uses external data? Thank you very much.

", "rawMarkdown": "May I ask if 0.800 uses external data? Thank you very much.", "votes": 1, "replies": [ { "id": 3179116, "postDate": "2025-04-15T04:11:32.910Z", "content": "

Thanks for the question.
\nI didn't use external data.

", "rawMarkdown": "Thanks for the question. \nI didn't use external data.", "votes": 4 } ] }, { "id": 3179304, "postDate": "2025-04-15T09:31:47.897Z", "content": "

Thank you for sharing such exciting exp result!
\nBy the way, how much does LB get boosted in post-processing?

", "rawMarkdown": "Thank you for sharing such exciting exp result!\nBy the way, how much does LB get boosted in post-processing?", "votes": 1, "replies": [ { "id": 3179354, "postDate": "2025-04-15T10:53:28.167Z", "content": "

Thank you. Actually, the post-processing score improvement is about 0.01 compared to the post-processing of the published notes.
\nThe main focus of the score improvement is on model training.

\n\n

contribute greatly to score improvement

\n

And I think the difficult part of this competition is that there is no regularity between the above devices and LB improvement!

", "rawMarkdown": "Thank you. Actually, the post-processing score improvement is about 0.01 compared to the post-processing of the published notes.\nThe main focus of the score improvement is on model training.\n- Batch size \n- Pre-processing of training data \n- Extraction method of training data \n\ncontribute greatly to score improvement \n\nAnd I think the difficult part of this competition is that there is no regularity between the above devices and LB improvement!", "votes": 8 }, { "id": 3179364, "postDate": "2025-04-15T10:59:07.347Z", "content": "

@minfuka what do you mean by extraction method of the training data, from what i've observed xy slices give the best performance, or do you mean something else?

", "rawMarkdown": "@minfuka what do you mean by extraction method of the training data, from what i've observed xy slices give the best performance, or do you mean something else?", "votes": 1 }, { "id": 3179387, "postDate": "2025-04-15T11:06:47.117Z", "content": "

I did not change the XY slice. This is the extraction method within XY slice. It is a decision of which slice to use as training data.

", "rawMarkdown": "I did not change the XY slice. This is the extraction method within XY slice. It is a decision of which slice to use as training data.", "votes": 3 }, { "id": 3180190, "postDate": "2025-04-16T09:16:00.673Z", "content": "

@minfuka \"And I think the difficult part of this competition is that there is no regularity between the above devices and LB improvement!\"

\n

imagine there is a score for each predicted motor in your local CV validation set.
\nnow you group your prediction boxes into bins:
\nbin0: score from 0. to 0.1
\nbin1: score from 0.1 to 0.2
\nbin2: score from 0.1 to 0.3
\n….

\n

assume that most of hidden test motor are in bin-n

\n

you can improve your local CV by improving e.g. only validation in bin-m.
\nbut this will not improve lb score at all.

\n

if you improve local cv of bin-n (same as hidden test), then lb would improve.

\n

in summary,

\n
    \n
  1. where is bin-n? i estimate hidden test is in about 0.45 to 0.65
    2. \n
  2. we need to find train samples in these range and improve them
    3. \n
", "rawMarkdown": "@minfuka \"And I think the difficult part of this competition is that there is no regularity between the above devices and LB improvement!\"\n\nimagine there is a score for each predicted motor in your local CV validation set.\nnow you group your prediction boxes into bins: \nbin0: score from 0. to 0.1\nbin1: score from 0.1 to 0.2\nbin2: score from 0.1 to 0.3\n....\n\nassume that most of hidden test motor are in bin-n\n\nyou can improve your local CV by improving e.g. only validation in bin-m.\nbut this will not improve lb score at all.\n\nif you improve local cv of bin-n (same as hidden test), then lb would improve. \n\nin summary,\n1. where is bin-n? i estimate hidden test is in about 0.45 to 0.65\n2. we need to find train samples in these range and improve them", "votes": 5 } ] }, { "id": 3179438, "postDate": "2025-04-15T11:54:37.413Z", "content": "

It is important to use larger image sizes for the input to YOLO😄

", "rawMarkdown": "It is important to use larger image sizes for the input to YOLO😄", "votes": 6 }, { "id": 3181054, "postDate": "2025-04-17T12:08:14.100Z", "content": "

May I ask which version of YOLO you are using?

", "rawMarkdown": "May I ask which version of YOLO you are using?", "replies": [ { "id": 3181093, "postDate": "2025-04-17T13:06:29.050Z", "content": "

I dare not say this, but it is true in my mind.

\n

＞Bigger models do not necessarily score better
\n＞Newer versions do not necessarily score better

\n

But from the reports I've heard from others, I don't think it's possible to say that a particular version is advantageous.
\nIn that sense, this competition is difficult.

", "rawMarkdown": "I dare not say this, but it is true in my mind.\n\n＞Bigger models do not necessarily score better\n＞Newer versions do not necessarily score better\n\nBut from the reports I've heard from others, I don't think it's possible to say that a particular version is advantageous.\nIn that sense, this competition is difficult.", "votes": 1 }, { "id": 3181100, "postDate": "2025-04-17T13:15:51.437Z", "content": "

Thank you for your reply.

", "rawMarkdown": "Thank you for your reply." } ] }, { "id": 3182073, "postDate": "2025-04-18T17:44:52.917Z", "content": "

@minfuka 0.8 on the leaderboard with a solo model is quite impressive!

\n

@hideyukizushi has achieved a 0.732 LB score shared publicly with a mAP of 0.948. I was wondering, would you be comfortable sharing the mAP of your model during training?

\n

It would really help a lot. I'm a newbie and still trying to understand how overfitting works. I assumed that if a model achieves a mAP of 0.97 or 0.98, it means it's overfitting? Because to reach that level, we'd usually have to train for 200 to 300 epochs, right?

", "rawMarkdown": "@minfuka `0.8` on the leaderboard with a `solo model` is quite impressive! \n\n@hideyukizushi has achieved a 0.732 LB score shared publicly with a `mAP of 0.948`. I was wondering, would you be comfortable sharing the mAP of your model during training? \n\nIt would really help a lot. I'm a newbie and still trying to understand how overfitting works. I assumed that if a model achieves a mAP of 0.97 or 0.98, it means it's overfitting? Because to reach that level, we'd usually have to train for 200 to 300 epochs, right?", "votes": 1, "replies": [ { "id": 3182219, "postDate": "2025-04-18T23:31:39.590Z", "content": "

Thanks for the good question

\n

I can't give you specific numbers, but in my experiments there is no relationship between improving mAP and improving LB score (this is one of the reasons why this competition is so difficult)

\n

Normally yolo training improves performance by increasing the number of epochs, but in this competition it is not always However, common sense does not necessarily apply in this competition. My guess is that there may be a discrepancy between the validationdata extracted from the training data and the competition testdata. @hengck23 gave a hint about this in the thread above.

", "rawMarkdown": "Thanks for the good question \n\nI can't give you specific numbers, but in my experiments there is no relationship between improving mAP and improving LB score (this is one of the reasons why this competition is so difficult) \n\nNormally yolo training improves performance by increasing the number of epochs, but in this competition it is not always However, common sense does not necessarily apply in this competition. My guess is that there may be a discrepancy between the validationdata extracted from the training data and the competition testdata. @hengck23 gave a hint about this in the thread above.\n", "votes": 1 } ] } ] }, { "id": 3178433, "postDate": "2025-04-14T07:11:00.343Z", "content": "

I was wondering whats the highest public LB score has YOLO achieved till now?

", "rawMarkdown": "I was wondering whats the highest public LB score has YOLO achieved till now?", "votes": 11 }, { "id": 3178694, "postDate": "2025-04-14T13:34:26.430Z", "content": "

It is hard to say the highest that specific model CAN achieve because everyone has different training hyperparameters, pre and post processing techniques, augmentations, and so on so it is hard to say what the highest you can get with YOLO is. I recommend comparing different model architectures under constant environments to compare such as hyperparameter tuning 3 different architectures and comparing their best CV performance to see what architecture works best here.

", "rawMarkdown": "It is hard to say the highest that specific model CAN achieve because everyone has different training hyperparameters, pre and post processing techniques, augmentations, and so on so it is hard to say what the highest you can get with YOLO is. I recommend comparing different model architectures under constant environments to compare such as hyperparameter tuning 3 different architectures and comparing their best CV performance to see what architecture works best here.", "votes": -1, "replies": [ { "id": 3178702, "postDate": "2025-04-14T13:48:36.970Z", "content": "

@connorjd pardon but i meant to say what is the highest LB score that yolo has achieved in this competition so far

", "rawMarkdown": "@connorjd pardon but i meant to say what is the highest LB score that yolo has achieved in this competition so far" } ] }, { "id": 3179713, "postDate": "2025-04-15T16:04:59.383Z", "content": "

Thanks for sharing my LB is with single model 0.61 best till now I am kind of stuk here.

", "rawMarkdown": "Thanks for sharing my LB is with single model 0.61 best till now I am kind of stuk here." }, { "id": 3178497, "postDate": "2025-04-14T08:58:10.567Z", "content": "

YOLO models often do very well, especially when speed is a priority.

\n

However, they sometimes don’t hit the very top of the LB, especially when very heavy, customized pipelines are used by others.

\n

If you're looking at a specific public leaderboard or dataset (COCO, Pascal VOC, or a Kaggle comp)

", "rawMarkdown": "YOLO models often do very well, especially when speed is a priority.\n\nHowever, they sometimes don’t hit the very top of the LB, especially when very heavy, customized pipelines are used by others.\n\nIf you're looking at a specific public leaderboard or dataset (COCO, Pascal VOC, or a Kaggle comp)" }, { "id": 3182083, "postDate": "2025-04-18T17:49:55.273Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3179012, "author_name": "min fuka", "author_url": "", "post_date": "2025-04-14T22:17:05.267000", "content": "

My LB0.800 is a single model of Yolo

\n\n

There is some ingenuity in post processing
\nBut I didn't use TTA.

", "votes": 17, "replies": [ { "id": 3179032, "author_name": "Andrei Zamfir", "author_url": "", "post_date": "2025-04-14T23:40:02.943000", "content": "

Impressive solo score, post processing is also something I found to require a lot of attention and creativity

", "votes": 1, "replies": [] }, { "id": 3179112, "author_name": "Handudu", "author_url": "", "post_date": "2025-04-15T03:56:31.227000", "content": "

May I ask if 0.800 uses external data? Thank you very much.

", "votes": 1, "replies": [ { "id": 3179116, "author_name": "min fuka", "author_url": "", "post_date": "2025-04-15T04:11:32.910000", "content": "

Thanks for the question.
\nI didn't use external data.

", "votes": 4, "replies": [] } ] }, { "id": 3179304, "author_name": "yukiZ", "author_url": "", "post_date": "2025-04-15T09:31:47.897000", "content": "

Thank you for sharing such exciting exp result!
\nBy the way, how much does LB get boosted in post-processing?

", "votes": 1, "replies": [ { "id": 3179354, "author_name": "min fuka", "author_url": "", "post_date": "2025-04-15T10:53:28.167000", "content": "

Thank you. Actually, the post-processing score improvement is about 0.01 compared to the post-processing of the published notes.
\nThe main focus of the score improvement is on model training.

\n\n

contribute greatly to score improvement

\n

And I think the difficult part of this competition is that there is no regularity between the above devices and LB improvement!

", "votes": 8, "replies": [] }, { "id": 3179364, "author_name": "m4nocha", "author_url": "", "post_date": "2025-04-15T10:59:07.347000", "content": "

@minfuka what do you mean by extraction method of the training data, from what i've observed xy slices give the best performance, or do you mean something else?

", "votes": 1, "replies": [] }, { "id": 3179387, "author_name": "min fuka", "author_url": "", "post_date": "2025-04-15T11:06:47.117000", "content": "

I did not change the XY slice. This is the extraction method within XY slice. It is a decision of which slice to use as training data.

", "votes": 3, "replies": [] }, { "id": 3180190, "author_name": "hengck23", "author_url": "", "post_date": "2025-04-16T09:16:00.673000", "content": "

@minfuka \"And I think the difficult part of this competition is that there is no regularity between the above devices and LB improvement!\"

\n

imagine there is a score for each predicted motor in your local CV validation set.
\nnow you group your prediction boxes into bins:
\nbin0: score from 0. to 0.1
\nbin1: score from 0.1 to 0.2
\nbin2: score from 0.1 to 0.3
\n….

\n

assume that most of hidden test motor are in bin-n

\n

you can improve your local CV by improving e.g. only validation in bin-m.
\nbut this will not improve lb score at all.

\n

if you improve local cv of bin-n (same as hidden test), then lb would improve.

\n

in summary,

\n
    \n
  1. where is bin-n? i estimate hidden test is in about 0.45 to 0.65
    2. \n
  2. we need to find train samples in these range and improve them
    3. \n
", "votes": 5, "replies": [] } ] }, { "id": 3179438, "author_name": "Bull", "author_url": "", "post_date": "2025-04-15T11:54:37.413000", "content": "

It is important to use larger image sizes for the input to YOLO😄

", "votes": 6, "replies": [] }, { "id": 3181054, "author_name": "hukaixin", "author_url": "", "post_date": "2025-04-17T12:08:14.100000", "content": "

May I ask which version of YOLO you are using?

", "votes": 0, "replies": [ { "id": 3181093, "author_name": "min fuka", "author_url": "", "post_date": "2025-04-17T13:06:29.050000", "content": "

I dare not say this, but it is true in my mind.

\n

＞Bigger models do not necessarily score better
\n＞Newer versions do not necessarily score better

\n

But from the reports I've heard from others, I don't think it's possible to say that a particular version is advantageous.
\nIn that sense, this competition is difficult.

", "votes": 1, "replies": [] }, { "id": 3181100, "author_name": "hukaixin", "author_url": "", "post_date": "2025-04-17T13:15:51.437000", "content": "

Thank you for your reply.

", "votes": 0, "replies": [] } ] }, { "id": 3182073, "author_name": "Sangram Patil", "author_url": "", "post_date": "2025-04-18T17:44:52.917000", "content": "

@minfuka 0.8 on the leaderboard with a solo model is quite impressive!

\n

@hideyukizushi has achieved a 0.732 LB score shared publicly with a mAP of 0.948. I was wondering, would you be comfortable sharing the mAP of your model during training?

\n

It would really help a lot. I'm a newbie and still trying to understand how overfitting works. I assumed that if a model achieves a mAP of 0.97 or 0.98, it means it's overfitting? Because to reach that level, we'd usually have to train for 200 to 300 epochs, right?

", "votes": 1, "replies": [ { "id": 3182219, "author_name": "min fuka", "author_url": "", "post_date": "2025-04-18T23:31:39.590000", "content": "

Thanks for the good question

\n

I can't give you specific numbers, but in my experiments there is no relationship between improving mAP and improving LB score (this is one of the reasons why this competition is so difficult)

\n

Normally yolo training improves performance by increasing the number of epochs, but in this competition it is not always However, common sense does not necessarily apply in this competition. My guess is that there may be a discrepancy between the validationdata extracted from the training data and the competition testdata. @hengck23 gave a hint about this in the thread above.

", "votes": 1, "replies": [] } ] } ] }, { "id": 3178694, "author_name": "Connor", "author_url": "", "post_date": "2025-04-14T13:34:26.430000", "content": "

It is hard to say the highest that specific model CAN achieve because everyone has different training hyperparameters, pre and post processing techniques, augmentations, and so on so it is hard to say what the highest you can get with YOLO is. I recommend comparing different model architectures under constant environments to compare such as hyperparameter tuning 3 different architectures and comparing their best CV performance to see what architecture works best here.

", "votes": -1, "replies": [ { "id": 3178702, "author_name": "m4nocha", "author_url": "", "post_date": "2025-04-14T13:48:36.970000", "content": "

@connorjd pardon but i meant to say what is the highest LB score that yolo has achieved in this competition so far

", "votes": 0, "replies": [] } ] }, { "id": 3179713, "author_name": "", "author_url": "", "post_date": "2025-04-15T16:04:59.383000", "content": "

Thanks for sharing my LB is with single model 0.61 best till now I am kind of stuk here.

", "votes": 0, "replies": [] }, { "id": 3178497, "author_name": "Ali Kabirzadeh", "author_url": "", "post_date": "2025-04-14T08:58:10.567000", "content": "

YOLO models often do very well, especially when speed is a priority.

\n

However, they sometimes don’t hit the very top of the LB, especially when very heavy, customized pipelines are used by others.

\n

If you're looking at a specific public leaderboard or dataset (COCO, Pascal VOC, or a Kaggle comp)

", "votes": 0, "replies": [] }, { "id": 3182083, "author_name": "", "author_url": "", "post_date": "2025-04-18T17:49:55.273000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3179012": "My LB0.800 is a single model of Yolo \n- Bigger models do not necessarily score better \n- Newer versions do not necessarily score better \nI know that\n\nThere is some ingenuity in post processing\nBut I didn't use TTA.", "3178433": "I was wondering whats the highest public LB score has YOLO achieved till now?", "3178694": "It is hard to say the highest that specific model CAN achieve because everyone has different training hyperparameters, pre and post processing techniques, augmentations, and so on so it is hard to say what the highest you can get with YOLO is. I recommend comparing different model architectures under constant environments to compare such as hyperparameter tuning 3 different architectures and comparing their best CV performance to see what architecture works best here.", "3179713": "Thanks for sharing my LB is with single model 0.61 best till now I am kind of stuk here.", "3178497": "YOLO models often do very well, especially when speed is a priority.\n\nHowever, they sometimes don’t hit the very top of the LB, especially when very heavy, customized pipelines are used by others.\n\nIf you're looking at a specific public leaderboard or dataset (COCO, Pascal VOC, or a Kaggle comp)", "3182083": "" } }