{ "id": 567234, "title": "object detection task vs segmentation task:", "url": "/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/567234", "author_name": "", "post_date": "2025-03-09T08:21:07.410412800Z", "votes": 3, "comment_count": 22, "views": 0, "content": "

there is many similarities in object detection and segmentation , i think this competition is about object detection not segmentation such our task to localize the motor by given their coordinate (axis) if there exist in a tomogram, i am not expert yet and i am not sure (this is may to be wrong) but this way in public notebook that are published yolo very outperformed other NNs architectures, correct me if i am wrong, what is your suggestions

", "messages": [ { "id": "3145025", "postDate": "03/09/2025 08:21:07", "content": "

there is many similarities in object detection and segmentation , i think this competition is about object detection not segmentation such our task to localize the motor by given their coordinate (axis) if there exist in a tomogram, i am not expert yet and i am not sure (this is may to be wrong) but this way in public notebook that are published yolo very outperformed other NNs architectures, correct me if i am wrong, what is your suggestions

", "rawMarkdown": "there is many similarities in object detection and segmentation , i think this competition is about object detection not segmentation such our task to localize the motor by given their coordinate (axis) if there exist in a tomogram, i am not expert yet and i am not sure (this is may to be wrong) but this way in public notebook that are published yolo very outperformed other NNs architectures, correct me if i am wrong, what is your suggestions", "votes": null }, { "id": "3145028", "postDate": "03/09/2025 08:34:37", "content": "

It's too early to tell which approach would work better. Both of them are worth trying. Segmentation might work well depending on your target transformation.

", "rawMarkdown": "It's too early to tell which approach would work better. Both of them are worth trying. Segmentation might work well depending on your target transformation.", "votes": null }, { "id": "3145044", "postDate": "03/09/2025 09:07:16", "content": "

@gunesevitan you're right , till now , i try many architectures (without any postprocessing, data augmentation , reconstruction , denoizing) and does not perform well

", "rawMarkdown": "gunesevitan you're right , till now , i try many architectures (without any postprocessing, data augmentation , reconstruction , denoizing) and does not perform well", "votes": null }, { "id": "3145733", "postDate": "03/10/2025 07:56:15", "content": "

I've seen someone create gaussian balls based on the target center points. Maybe it will be great to use it.

", "rawMarkdown": "I've seen someone create gaussian balls based on the target center points. Maybe it will be great to use it.", "votes": null }, { "id": "3145736", "postDate": "03/10/2025 08:00:22", "content": "

@tom99763 it hard to know what it should work and not work unless you experiment it

", "rawMarkdown": "tom99763 it hard to know what it should work and not work unless you experiment it", "votes": null }, { "id": "3145739", "postDate": "03/10/2025 08:08:47", "content": "

@saidkoussi Yeah I think I can test the power of RTX 5090. I just bought it yesterday.

", "rawMarkdown": "saidkoussi Yeah I think I can test the power of RTX 5090. I just bought it yesterday.", "votes": null }, { "id": "3145743", "postDate": "03/10/2025 08:14:13", "content": "

@tom99763 great you are on the right track , my local machine is RTX 5070,(12 GB VRAM) , it struggle with large CNN architecture , but i think , the skills is all we need

", "rawMarkdown": "tom99763 great you are on the right track , my local machine is RTX 5070,(12 GB VRAM) , it struggle with large CNN architecture , but i think , the skills is all we need", "votes": null }, { "id": "3145837", "postDate": "03/10/2025 09:48:31", "content": "

@saidkoussi how large? I currently run 3D UNet it works well.

", "rawMarkdown": "saidkoussi how large? I currently run 3D UNet it works well.", "votes": null }, { "id": "3145974", "postDate": "03/10/2025 12:30:36", "content": "

@tom99763 yes 3D UNet, but if you combine it with pretrained model such ResNet the Out Of Memory arise , if you don't mind do you use yolo in the training or not

", "rawMarkdown": "tom99763 yes 3D UNet, but if you combine it with pretrained model such ResNet the Out Of Memory arise , if you don't mind do you use yolo in the training or not", "votes": null }, { "id": "3145977", "postDate": "03/10/2025 12:36:46", "content": "

Do you set requires grad=false?

", "rawMarkdown": "Do you set requires grad=false?", "votes": null }, { "id": "3145983", "postDate": "03/10/2025 12:43:23", "content": "

not i try this :

\n
= \nimport os\nos.environ[] = \n\n= torch.cuda.amp.GradScaler() \n        with torch.cuda.amp.autocast():          \n\n                     \n                 \n
\n

such the model perfoming very poorely , so i think i must think about the target transformation , preaparing datast , not about CNN architecture , i think the good approach is to make solide pipeline for promesing score and choose good NNs architecure and try enhance it.

", "rawMarkdown": "not i try this :\n``` \nbatch_size = 1\nimport os\nos.environ[\"PYTORCH_CUDA_ALLOC_CONF\"] = \"expandable_segments:True\"\n# during training:\nscaler = torch.cuda.amp.GradScaler() \n with torch.cuda.amp.autocast(): \n\n scaler.scale(loss).backward() \n scaler.step(optimizer) \n```\nsuch the model perfoming very poorely , so i think i must think about the target transformation , preaparing datast , not about CNN architecture , i think the good approach is to make solide pipeline for promesing score and choose good NNs architecure and try enhance it.", "votes": null }, { "id": "3145985", "postDate": "03/10/2025 12:49:36", "content": "

Did you use class-weight to address very sparse segmentation mask?

", "rawMarkdown": "Did you use class-weight to address very sparse segmentation mask?", "votes": null }, { "id": "3146002", "postDate": "03/10/2025 13:08:42", "content": "

no , i want to take approach of object detection task not segmentation approach , but i'll try both

", "rawMarkdown": "no , i want to take approach of object detection task not segmentation approach , but i'll try both", "votes": null }, { "id": "3146077", "postDate": "03/10/2025 13:53:36", "content": "

both and ensemble is best ;)

\n

at least it was best in last object detection tomogram competition

", "rawMarkdown": "both and ensemble is best ;)\n\nat least it was best in last object detection tomogram competition", "votes": null }, { "id": "3146080", "postDate": "03/10/2025 13:56:28", "content": "

@christofhenkel hi Dieter I just finish transferring your github code from cryto comp into here and you just jump in.

", "rawMarkdown": "christofhenkel hi Dieter I just finish transferring your github code from cryto comp into here and you just jump in.", "votes": null }, { "id": "3146084", "postDate": "03/10/2025 13:57:31", "content": "

might join after AIMO2 competition

", "rawMarkdown": "might join after AIMO2 competition", "votes": null }, { "id": "3146089", "postDate": "03/10/2025 14:01:53", "content": "

absolutely, experimentation is all we need

", "rawMarkdown": "absolutely, experimentation is all we need", "votes": null }, { "id": "3146118", "postDate": "03/10/2025 14:35:00", "content": "

I think you're right. Since I plan to use the 3D U-Net, just like in the previous CZII competition. However, I've tried two or three methods to load data in the dataset class, and the results were really disappointing. Either there wasn't enough memory, or the data loading speed was extremely slow. Next, I want to give CNN+LSTM a try.

", "rawMarkdown": "I think you're right. Since I plan to use the 3D U-Net, just like in the previous CZII competition. However, I've tried two or three methods to load data in the dataset class, and the results were really disappointing. Either there wasn't enough memory, or the data loading speed was extremely slow. Next, I want to give CNN+LSTM a try.", "votes": null }, { "id": "3146120", "postDate": "03/10/2025 14:35:11", "content": "

@saidkoussi I was thinking about creating irregular edges based on the center points they provided. There's a lot of thing we need to try.

", "rawMarkdown": "saidkoussi I was thinking about creating irregular edges based on the center points they provided. There's a lot of thing we need to try.", "votes": null }, { "id": "3146133", "postDate": "03/10/2025 14:43:52", "content": "

yes , target transformation as @gunesevitan mentionned early , i think about creating bound boxing right now, from centers

", "rawMarkdown": "yes , target transformation as @gunesevitan mentionned early , i think about creating bound boxing right now, from centers", "votes": null }, { "id": "3146137", "postDate": "03/10/2025 14:52:02", "content": "

@switch9527 i'm also partipating in CZII competition , i stick with YOLO models , use all pretrained models available, fine tuning the hyperparameters … , but in my opinion , if the models does not yield promising results , we must try other approach not changing the models but the whole pepeline from data preparation to the inference

", "rawMarkdown": "switch9527 i'm also partipating in CZII competition , i stick with YOLO models , use all pretrained models available, fine tuning the hyperparameters ... , but in my opinion , if the models does not yield promising results , we must try other approach not changing the models but the whole pepeline from data preparation to the inference", "votes": null }, { "id": "3146148", "postDate": "03/10/2025 15:00:36", "content": "

I also try some crazy thing like point cloud, waiting to see the results.

", "rawMarkdown": "I also try some crazy thing like point cloud, waiting to see the results.", "votes": null }, { "id": "3156544", "postDate": "03/22/2025 08:59:13", "content": "

@christofhenkel Bro, past records are crazy! Bro, stop!💀

", "rawMarkdown": "christofhenkel Bro, past records are crazy! Bro, stop!💀", "votes": null } ], "comments": [ { "id": 3145028, "author_name": "gunesevitan", "author_url": "", "post_date": "03/09/2025 08:34:37", "content": "

It's too early to tell which approach would work better. Both of them are worth trying. Segmentation might work well depending on your target transformation.

", "votes": null, "replies": [ { "id": 3145044, "author_name": "saidkoussi", "author_url": "", "post_date": "03/09/2025 09:07:16", "content": "

@gunesevitan you're right , till now , i try many architectures (without any postprocessing, data augmentation , reconstruction , denoizing) and does not perform well

", "votes": null, "replies": [] }, { "id": 3145733, "author_name": "tom99763", "author_url": "", "post_date": "03/10/2025 07:56:15", "content": "

I've seen someone create gaussian balls based on the target center points. Maybe it will be great to use it.

", "votes": null, "replies": [ { "id": 3145736, "author_name": "saidkoussi", "author_url": "", "post_date": "03/10/2025 08:00:22", "content": "

@tom99763 it hard to know what it should work and not work unless you experiment it

", "votes": null, "replies": [ { "id": 3145739, "author_name": "tom99763", "author_url": "", "post_date": "03/10/2025 08:08:47", "content": "

@saidkoussi Yeah I think I can test the power of RTX 5090. I just bought it yesterday.

", "votes": null, "replies": [ { "id": 3145743, "author_name": "saidkoussi", "author_url": "", "post_date": "03/10/2025 08:14:13", "content": "

@tom99763 great you are on the right track , my local machine is RTX 5070,(12 GB VRAM) , it struggle with large CNN architecture , but i think , the skills is all we need

", "votes": null, "replies": [ { "id": 3145837, "author_name": "tom99763", "author_url": "", "post_date": "03/10/2025 09:48:31", "content": "

@saidkoussi how large? I currently run 3D UNet it works well.

", "votes": null, "replies": [ { "id": 3145974, "author_name": "saidkoussi", "author_url": "", "post_date": "03/10/2025 12:30:36", "content": "

@tom99763 yes 3D UNet, but if you combine it with pretrained model such ResNet the Out Of Memory arise , if you don't mind do you use yolo in the training or not

", "votes": null, "replies": [ { "id": 3145977, "author_name": "tom99763", "author_url": "", "post_date": "03/10/2025 12:36:46", "content": "

Do you set requires grad=false?

", "votes": null, "replies": [ { "id": 3145983, "author_name": "saidkoussi", "author_url": "", "post_date": "03/10/2025 12:43:23", "content": "

not i try this :

\n
= \nimport os\nos.environ[] = \n\n= torch.cuda.amp.GradScaler() \n        with torch.cuda.amp.autocast():          \n\n                     \n                 \n
\n

such the model perfoming very poorely , so i think i must think about the target transformation , preaparing datast , not about CNN architecture , i think the good approach is to make solide pipeline for promesing score and choose good NNs architecure and try enhance it.

", "votes": null, "replies": [ { "id": 3145985, "author_name": "tom99763", "author_url": "", "post_date": "03/10/2025 12:49:36", "content": "

Did you use class-weight to address very sparse segmentation mask?

", "votes": null, "replies": [ { "id": 3146002, "author_name": "saidkoussi", "author_url": "", "post_date": "03/10/2025 13:08:42", "content": "

no , i want to take approach of object detection task not segmentation approach , but i'll try both

", "votes": null, "replies": [] } ] } ] } ] } ] } ] } ] } ] } ] } ] } ] }, { "id": 3146077, "author_name": "christofhenkel", "author_url": "", "post_date": "03/10/2025 13:53:36", "content": "

both and ensemble is best ;)

\n

at least it was best in last object detection tomogram competition

", "votes": null, "replies": [ { "id": 3146080, "author_name": "tom99763", "author_url": "", "post_date": "03/10/2025 13:56:28", "content": "

@christofhenkel hi Dieter I just finish transferring your github code from cryto comp into here and you just jump in.

", "votes": null, "replies": [ { "id": 3146084, "author_name": "christofhenkel", "author_url": "", "post_date": "03/10/2025 13:57:31", "content": "

might join after AIMO2 competition

", "votes": null, "replies": [ { "id": 3156544, "author_name": "sangrampatil5150", "author_url": "", "post_date": "03/22/2025 08:59:13", "content": "

@christofhenkel Bro, past records are crazy! Bro, stop!💀

", "votes": null, "replies": [] } ] } ] }, { "id": 3146089, "author_name": "saidkoussi", "author_url": "", "post_date": "03/10/2025 14:01:53", "content": "

absolutely, experimentation is all we need

", "votes": null, "replies": [ { "id": 3146120, "author_name": "tom99763", "author_url": "", "post_date": "03/10/2025 14:35:11", "content": "

@saidkoussi I was thinking about creating irregular edges based on the center points they provided. There's a lot of thing we need to try.

", "votes": null, "replies": [ { "id": 3146133, "author_name": "saidkoussi", "author_url": "", "post_date": "03/10/2025 14:43:52", "content": "

yes , target transformation as @gunesevitan mentionned early , i think about creating bound boxing right now, from centers

", "votes": null, "replies": [ { "id": 3146148, "author_name": "tom99763", "author_url": "", "post_date": "03/10/2025 15:00:36", "content": "

I also try some crazy thing like point cloud, waiting to see the results.

", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 3146118, "author_name": "switch9527", "author_url": "", "post_date": "03/10/2025 14:35:00", "content": "

I think you're right. Since I plan to use the 3D U-Net, just like in the previous CZII competition. However, I've tried two or three methods to load data in the dataset class, and the results were really disappointing. Either there wasn't enough memory, or the data loading speed was extremely slow. Next, I want to give CNN+LSTM a try.

", "votes": null, "replies": [ { "id": 3146137, "author_name": "saidkoussi", "author_url": "", "post_date": "03/10/2025 14:52:02", "content": "

@switch9527 i'm also partipating in CZII competition , i stick with YOLO models , use all pretrained models available, fine tuning the hyperparameters … , but in my opinion , if the models does not yield promising results , we must try other approach not changing the models but the whole pepeline from data preparation to the inference

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3145025": "there is many similarities in object detection and segmentation , i think this competition is about object detection not segmentation such our task to localize the motor by given their coordinate (axis) if there exist in a tomogram, i am not expert yet and i am not sure (this is may to be wrong) but this way in public notebook that are published yolo very outperformed other NNs architectures, correct me if i am wrong, what is your suggestions", "3145028": "It's too early to tell which approach would work better. Both of them are worth trying. Segmentation might work well depending on your target transformation.", "3145044": "gunesevitan you're right , till now , i try many architectures (without any postprocessing, data augmentation , reconstruction , denoizing) and does not perform well", "3145733": "I've seen someone create gaussian balls based on the target center points. Maybe it will be great to use it.", "3145736": "tom99763 it hard to know what it should work and not work unless you experiment it", "3145739": "saidkoussi Yeah I think I can test the power of RTX 5090. I just bought it yesterday.", "3145743": "tom99763 great you are on the right track , my local machine is RTX 5070,(12 GB VRAM) , it struggle with large CNN architecture , but i think , the skills is all we need", "3145837": "saidkoussi how large? I currently run 3D UNet it works well.", "3145974": "tom99763 yes 3D UNet, but if you combine it with pretrained model such ResNet the Out Of Memory arise , if you don't mind do you use yolo in the training or not", "3145977": "Do you set requires grad=false?", "3145983": "not i try this :\n``` \nbatch_size = 1\nimport os\nos.environ[\"PYTORCH_CUDA_ALLOC_CONF\"] = \"expandable_segments:True\"\n# during training:\nscaler = torch.cuda.amp.GradScaler() \n with torch.cuda.amp.autocast(): \n\n scaler.scale(loss).backward() \n scaler.step(optimizer) \n```\nsuch the model perfoming very poorely , so i think i must think about the target transformation , preaparing datast , not about CNN architecture , i think the good approach is to make solide pipeline for promesing score and choose good NNs architecure and try enhance it.", "3145985": "Did you use class-weight to address very sparse segmentation mask?", "3146002": "no , i want to take approach of object detection task not segmentation approach , but i'll try both", "3146077": "both and ensemble is best ;)\n\nat least it was best in last object detection tomogram competition", "3146080": "christofhenkel hi Dieter I just finish transferring your github code from cryto comp into here and you just jump in.", "3146084": "might join after AIMO2 competition", "3146089": "absolutely, experimentation is all we need", "3146118": "I think you're right. Since I plan to use the 3D U-Net, just like in the previous CZII competition. However, I've tried two or three methods to load data in the dataset class, and the results were really disappointing. Either there wasn't enough memory, or the data loading speed was extremely slow. Next, I want to give CNN+LSTM a try.", "3146120": "saidkoussi I was thinking about creating irregular edges based on the center points they provided. There's a lot of thing we need to try.", "3146133": "yes , target transformation as @gunesevitan mentionned early , i think about creating bound boxing right now, from centers", "3146137": "switch9527 i'm also partipating in CZII competition , i stick with YOLO models , use all pretrained models available, fine tuning the hyperparameters ... , but in my opinion , if the models does not yield promising results , we must try other approach not changing the models but the whole pepeline from data preparation to the inference", "3146148": "I also try some crazy thing like point cloud, waiting to see the results.", "3156544": "christofhenkel Bro, past records are crazy! Bro, stop!💀" }, "source": "meta" }