{ "id": 579247, "title": "What is the key point to reach 0.8?", "url": "/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/579247", "author_name": "Seeing Times", "post_date": "2025-05-16T07:20:10.401000", "votes": 6, "comment_count": 35, "views": 0, "content": "

I'm currently using a YOLO model and have achieved an LB score of 0.776 without external data, TTA, or heavy data augmentation. What can I do to improve my score to 0.8 or higher? Any recommendations?

", "messages": [ { "id": 3202990, "postDate": "2025-05-16T07:20:10.400Z", "content": "

I'm currently using a YOLO model and have achieved an LB score of 0.776 without external data, TTA, or heavy data augmentation. What can I do to improve my score to 0.8 or higher? Any recommendations?

", "rawMarkdown": "I'm currently using a YOLO model and have achieved an LB score of 0.776 without external data, TTA, or heavy data augmentation. What can I do to improve my score to 0.8 or higher? Any recommendations?", "votes": 5 }, { "id": 3207703, "postDate": "2025-05-23T06:38:42.227Z", "content": "

Thanks for discussions, give me some hints.

", "rawMarkdown": "Thanks for discussions, give me some hints.", "votes": 1 }, { "id": 3203860, "postDate": "2025-05-17T12:00:32.897Z", "content": "

I got 75.5 using YOLO11, but with some post-processing technique I I wa it up to around 80. If anyone has a single model, doesn't have to be yolo, that performs well and wants to team up, let me know. We might be able to improve it by a couple of points.

", "rawMarkdown": "I got 75.5 using YOLO11, but with some post-processing technique I I wa it up to around 80. If anyone has a single model, doesn't have to be yolo, that performs well and wants to team up, let me know. We might be able to improve it by a couple of points. ", "votes": 1, "replies": [ { "id": 3203892, "postDate": "2025-05-17T12:41:48.523Z", "content": "

I got a 2d convnext model with 0.73 score😂

", "rawMarkdown": "I got a 2d convnext model with 0.73 score😂" }, { "id": 3204755, "postDate": "2025-05-18T18:28:35.490Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3204759, "postDate": "2025-05-18T18:34:42.693Z", "content": "

If your ok to try out 2d+3d / 2.5 d / Equivariant (I have tried this before its amazing considering the ''domain shift'' )approaches , then we can distribute our work and experiment …..

", "rawMarkdown": "If your ok to try out 2d+3d / 2.5 d / Equivariant (I have tried this before its amazing considering the ''domain shift'' )approaches , then we can distribute our work and experiment .....", "isDeleted": true }, { "id": 3206217, "postDate": "2025-05-21T03:31:14.653Z", "content": "

Hi, do you have strong strategy in post-processing? We have decent single model without pp, would you like to share more? Are you still looking for teammate? Let's see if we can team up.

", "rawMarkdown": "Hi, do you have strong strategy in post-processing? We have decent single model without pp, would you like to share more? Are you still looking for teammate? Let's see if we can team up.", "replies": [ { "id": 3206701, "postDate": "2025-05-21T17:03:32.693Z", "content": "

I dont think I'll consider it strong , its just basic observation , but my next plan for post processing is whole lot different , training a new model

", "rawMarkdown": "I dont think I'll consider it strong , its just basic observation , but my next plan for post processing is whole lot different , training a new model ", "isDeleted": true } ] } ] }, { "id": 3203855, "postDate": "2025-05-17T11:54:24.293Z", "content": "

@seeingtimes Basically, changing yolo8 to yolo11 can easily get 0.8.

", "rawMarkdown": "@seeingtimes Basically, changing yolo8 to yolo11 can easily get 0.8. ", "votes": 1, "replies": [ { "id": 3203857, "postDate": "2025-05-17T11:56:18.137Z", "content": "

Yolo 11 L？

", "rawMarkdown": "Yolo 11 L？", "votes": 1, "replies": [ { "id": 3203861, "postDate": "2025-05-17T12:03:19.017Z", "content": "

yolo11s is enough

", "rawMarkdown": "yolo11s is enough" }, { "id": 3203862, "postDate": "2025-05-17T12:03:49.760Z", "content": "

thx, I gonna try now

", "rawMarkdown": "thx, I gonna try now" }, { "id": 3204246, "postDate": "2025-05-18T02:55:57.587Z", "content": "

got 0.047 with yolo11s (same setting as yolov8)😶‍🌫️

", "rawMarkdown": "got 0.047 with yolo11s (same setting as yolov8)😶‍🌫️" } ] } ] }, { "id": 3203126, "postDate": "2025-05-16T10:32:05.707Z", "content": "

By using mhaf-yolo model mentioned in the public notebook and adjusting the threshold, I can get 0.803

", "rawMarkdown": "By using mhaf-yolo model mentioned in the public notebook and adjusting the threshold, I can get 0.803", "votes": 1, "replies": [ { "id": 3203164, "postDate": "2025-05-16T11:26:16.110Z", "content": "

I'm using YOLOv8, as mentioned in the discussion, but I just can't reach 0.8. There must be something I haven't done yet.

", "rawMarkdown": "I'm using YOLOv8, as mentioned in the discussion, but I just can't reach 0.8. There must be something I haven't done yet." }, { "id": 3203587, "postDate": "2025-05-17T01:15:35.273Z", "content": "

did u just used the weights , or trained at your end ??

\n

How did your guys parse data look like ??

\n

Did you just used the starter version or edited it??

", "rawMarkdown": "did u just used the weights , or trained at your end ??\n\nHow did your guys parse data look like ??\n\nDid you just used the starter version or edited it??", "isDeleted": true, "replies": [ { "id": 3203590, "postDate": "2025-05-17T01:27:07.880Z", "content": "

Yes, I used public process code with tomo > 0, and I wrote my own train code(just for easily modify hyper parameters)

", "rawMarkdown": "Yes, I used public process code with tomo > 0, and I wrote my own train code(just for easily modify hyper parameters)" }, { "id": 3204751, "postDate": "2025-05-18T18:26:43.380Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3204758, "postDate": "2025-05-18T18:32:38.340Z", "content": "

That has some errors , (it labels only one motor per slice (even with multiple ones ) atleast in the last version I checked , hengck pointed it out in the comments ,

", "rawMarkdown": "That has some errors , (it labels only one motor per slice (even with multiple ones ) atleast in the last version I checked , hengck pointed it out in the comments ,", "isDeleted": true } ] }, { "id": 3204915, "postDate": "2025-05-19T03:41:10.577Z", "content": "

Hi @i2nfinit3y , I’m also trying to determine the optimal confidence threshold. Could you let me know whether you lowered the threshold to achieve the 0.803 score?

", "rawMarkdown": "Hi @i2nfinit3y , I’m also trying to determine the optimal confidence threshold. Could you let me know whether you lowered the threshold to achieve the 0.803 score?", "replies": [ { "id": 3205068, "postDate": "2025-05-19T09:08:30.873Z", "content": "

No, I raise the threshold.

", "rawMarkdown": "No, I raise the threshold.", "votes": 1 }, { "id": 3207056, "postDate": "2025-05-22T06:53:56.217Z", "content": "

What is your image size for YOLO model?

", "rawMarkdown": "What is your image size for YOLO model?" } ] } ] }, { "id": 3203085, "postDate": "2025-05-16T09:45:17.543Z", "content": "

I'm also curious. I used SAHI and barely reached 0.8~

", "rawMarkdown": "I'm also curious. I used SAHI and barely reached 0.8~", "votes": 1, "replies": [ { "id": 3203101, "postDate": "2025-05-16T10:03:13.380Z", "content": "

thx, I gonna try SAHI later😀

", "rawMarkdown": "thx, I gonna try SAHI later😀" }, { "id": 3203106, "postDate": "2025-05-16T10:09:07.967Z", "content": "

how did you train your model ?

", "rawMarkdown": "how did you train your model ?", "replies": [ { "id": 3203188, "postDate": "2025-05-16T12:05:57.697Z", "content": "

Just use normal augmentation, my single model without SAHI is also around 0.77.

", "rawMarkdown": "Just use normal augmentation, my single model without SAHI is also around 0.77." }, { "id": 3203195, "postDate": "2025-05-16T12:12:28.493Z", "content": "

Could you post your prediction on three dummy data?

", "rawMarkdown": "Could you post your prediction on three dummy data?" }, { "id": 3203208, "postDate": "2025-05-16T12:35:06.450Z", "content": "

maybe the data sahi is giving to the model is not the data the model has been trained with @heelerdeer

", "rawMarkdown": "maybe the data sahi is giving to the model is not the data the model has been trained with @heelerdeer " }, { "id": 3203586, "postDate": "2025-05-17T01:11:46.407Z", "content": "

could you tell me how did your parse data notebook look like - did u corrected anything or just used that starter version.

", "rawMarkdown": "could you tell me how did your parse data notebook look like - did u corrected anything or just used that starter version.", "isDeleted": true }, { "id": 3203591, "postDate": "2025-05-17T01:30:31.137Z", "content": "

just the stater version for prepare data

", "rawMarkdown": "just the stater version for prepare data" }, { "id": 3203603, "postDate": "2025-05-17T01:57:54.643Z", "content": "

I think I need to try more epochs (20-30 aint enough) what about you what are your epochs??

", "rawMarkdown": "I think I need to try more epochs (20-30 aint enough) what about you what are your epochs??", "isDeleted": true }, { "id": 3203611, "postDate": "2025-05-17T02:25:36.540Z", "content": "

50epochs，but the best pt file is always around 30

", "rawMarkdown": "50epochs，but the best pt file is always around 30" }, { "id": 3203661, "postDate": "2025-05-17T05:37:16.753Z", "content": "

hmm
\nand lr?
\nany scheduler??

", "rawMarkdown": "hmm \nand lr? \nany scheduler??", "isDeleted": true }, { "id": 3206917, "postDate": "2025-05-22T00:00:48.107Z", "content": "

le-4 with adamw what is yours？

", "rawMarkdown": "le-4 with adamw what is yours？" }, { "id": 3207603, "postDate": "2025-05-23T03:03:37.727Z", "content": "

My current LB is based on public weights of MHAF with post proc

", "rawMarkdown": "My current LB is based on public weights of MHAF with post proc", "isDeleted": true } ] }, { "id": 3203592, "postDate": "2025-05-17T01:30:57.173Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 3207703, "author_name": "", "author_url": "", "post_date": "2025-05-23T06:38:42.227000", "content": "

Thanks for discussions, give me some hints.

", "votes": 1, "replies": [] }, { "id": 3203860, "author_name": "GoMBap", "author_url": "", "post_date": "2025-05-17T12:00:32.897000", "content": "

I got 75.5 using YOLO11, but with some post-processing technique I I wa it up to around 80. If anyone has a single model, doesn't have to be yolo, that performs well and wants to team up, let me know. We might be able to improve it by a couple of points.

", "votes": 1, "replies": [ { "id": 3203892, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-17T12:41:48.523000", "content": "

I got a 2d convnext model with 0.73 score😂

", "votes": 0, "replies": [] }, { "id": 3204755, "author_name": "", "author_url": "", "post_date": "2025-05-18T18:28:35.490000", "content": "", "votes": 0, "replies": [] }, { "id": 3204759, "author_name": "", "author_url": "", "post_date": "2025-05-18T18:34:42.693000", "content": "

If your ok to try out 2d+3d / 2.5 d / Equivariant (I have tried this before its amazing considering the ''domain shift'' )approaches , then we can distribute our work and experiment …..

", "votes": 0, "replies": [] }, { "id": 3206217, "author_name": "shanzhong8", "author_url": "", "post_date": "2025-05-21T03:31:14.653000", "content": "

Hi, do you have strong strategy in post-processing? We have decent single model without pp, would you like to share more? Are you still looking for teammate? Let's see if we can team up.

", "votes": 0, "replies": [ { "id": 3206701, "author_name": "", "author_url": "", "post_date": "2025-05-21T17:03:32.693000", "content": "

I dont think I'll consider it strong , its just basic observation , but my next plan for post processing is whole lot different , training a new model

", "votes": 0, "replies": [] } ] } ] }, { "id": 3203855, "author_name": "Tom", "author_url": "", "post_date": "2025-05-17T11:54:24.293000", "content": "

@seeingtimes Basically, changing yolo8 to yolo11 can easily get 0.8.

", "votes": 1, "replies": [ { "id": 3203857, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-17T11:56:18.137000", "content": "

Yolo 11 L？

", "votes": 1, "replies": [ { "id": 3203861, "author_name": "Tom", "author_url": "", "post_date": "2025-05-17T12:03:19.017000", "content": "

yolo11s is enough

", "votes": 0, "replies": [] }, { "id": 3203862, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-17T12:03:49.760000", "content": "

thx, I gonna try now

", "votes": 0, "replies": [] }, { "id": 3204246, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-18T02:55:57.587000", "content": "

got 0.047 with yolo11s (same setting as yolov8)😶‍🌫️

", "votes": 0, "replies": [] } ] } ] }, { "id": 3203126, "author_name": "I2nfinit3y", "author_url": "", "post_date": "2025-05-16T10:32:05.707000", "content": "

By using mhaf-yolo model mentioned in the public notebook and adjusting the threshold, I can get 0.803

", "votes": 1, "replies": [ { "id": 3203164, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-16T11:26:16.110000", "content": "

I'm using YOLOv8, as mentioned in the discussion, but I just can't reach 0.8. There must be something I haven't done yet.

", "votes": 0, "replies": [] }, { "id": 3203587, "author_name": "", "author_url": "", "post_date": "2025-05-17T01:15:35.273000", "content": "

did u just used the weights , or trained at your end ??

\n

How did your guys parse data look like ??

\n

Did you just used the starter version or edited it??

", "votes": 0, "replies": [ { "id": 3203590, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-17T01:27:07.880000", "content": "

Yes, I used public process code with tomo > 0, and I wrote my own train code(just for easily modify hyper parameters)

", "votes": 0, "replies": [] }, { "id": 3204751, "author_name": "", "author_url": "", "post_date": "2025-05-18T18:26:43.380000", "content": "", "votes": 0, "replies": [] }, { "id": 3204758, "author_name": "", "author_url": "", "post_date": "2025-05-18T18:32:38.340000", "content": "

That has some errors , (it labels only one motor per slice (even with multiple ones ) atleast in the last version I checked , hengck pointed it out in the comments ,

", "votes": 0, "replies": [] } ] }, { "id": 3204915, "author_name": "Gao Xin", "author_url": "", "post_date": "2025-05-19T03:41:10.577000", "content": "

Hi @i2nfinit3y , I’m also trying to determine the optimal confidence threshold. Could you let me know whether you lowered the threshold to achieve the 0.803 score?

", "votes": 0, "replies": [ { "id": 3205068, "author_name": "I2nfinit3y", "author_url": "", "post_date": "2025-05-19T09:08:30.873000", "content": "

No, I raise the threshold.

", "votes": 1, "replies": [] }, { "id": 3207056, "author_name": "Rustam Bazarbayev", "author_url": "", "post_date": "2025-05-22T06:53:56.217000", "content": "

What is your image size for YOLO model?

", "votes": 0, "replies": [] } ] } ] }, { "id": 3203085, "author_name": "Heeler-Deer", "author_url": "", "post_date": "2025-05-16T09:45:17.543000", "content": "

I'm also curious. I used SAHI and barely reached 0.8~

", "votes": 1, "replies": [ { "id": 3203101, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-16T10:03:13.380000", "content": "

thx, I gonna try SAHI later😀

", "votes": 0, "replies": [] }, { "id": 3203106, "author_name": "Victor", "author_url": "", "post_date": "2025-05-16T10:09:07.967000", "content": "

how did you train your model ?

", "votes": 0, "replies": [ { "id": 3203188, "author_name": "Heeler-Deer", "author_url": "", "post_date": "2025-05-16T12:05:57.697000", "content": "

Just use normal augmentation, my single model without SAHI is also around 0.77.

", "votes": 0, "replies": [] }, { "id": 3203195, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-16T12:12:28.493000", "content": "

Could you post your prediction on three dummy data?

", "votes": 0, "replies": [] }, { "id": 3203208, "author_name": "Victor", "author_url": "", "post_date": "2025-05-16T12:35:06.450000", "content": "

maybe the data sahi is giving to the model is not the data the model has been trained with @heelerdeer

", "votes": 0, "replies": [] }, { "id": 3203586, "author_name": "", "author_url": "", "post_date": "2025-05-17T01:11:46.407000", "content": "

could you tell me how did your parse data notebook look like - did u corrected anything or just used that starter version.

", "votes": 0, "replies": [] }, { "id": 3203591, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-17T01:30:31.137000", "content": "

just the stater version for prepare data

", "votes": 0, "replies": [] }, { "id": 3203603, "author_name": "", "author_url": "", "post_date": "2025-05-17T01:57:54.643000", "content": "

I think I need to try more epochs (20-30 aint enough) what about you what are your epochs??

", "votes": 0, "replies": [] }, { "id": 3203611, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-17T02:25:36.540000", "content": "

50epochs，but the best pt file is always around 30

", "votes": 0, "replies": [] }, { "id": 3203661, "author_name": "", "author_url": "", "post_date": "2025-05-17T05:37:16.753000", "content": "

hmm
\nand lr?
\nany scheduler??

", "votes": 0, "replies": [] }, { "id": 3206917, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-05-22T00:00:48.107000", "content": "

le-4 with adamw what is yours？

", "votes": 0, "replies": [] }, { "id": 3207603, "author_name": "", "author_url": "", "post_date": "2025-05-23T03:03:37.727000", "content": "

My current LB is based on public weights of MHAF with post proc

", "votes": 0, "replies": [] } ] }, { "id": 3203592, "author_name": "", "author_url": "", "post_date": "2025-05-17T01:30:57.173000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "3202990": "I'm currently using a YOLO model and have achieved an LB score of 0.776 without external data, TTA, or heavy data augmentation. What can I do to improve my score to 0.8 or higher? Any recommendations?", "3207703": "Thanks for discussions, give me some hints.", "3203860": "I got 75.5 using YOLO11, but with some post-processing technique I I wa it up to around 80. If anyone has a single model, doesn't have to be yolo, that performs well and wants to team up, let me know. We might be able to improve it by a couple of points. ", "3203855": "@seeingtimes Basically, changing yolo8 to yolo11 can easily get 0.8. ", "3203126": "By using mhaf-yolo model mentioned in the public notebook and adjusting the threshold, I can get 0.803", "3203085": "I'm also curious. I used SAHI and barely reached 0.8~" } }