\n
![\"\"](\"https://www.kaggleusercontent.com/kf/205937086/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..UN465CQ9InL2lXiDGg5C_g.Bli54dxbWCbtrvgJb0_F-l9jpBhGnt_ED-yYTaE87Tae73D2jtCOko_8lnPEZSoQQSK2uGKcnqXMWImk-1TOgBJxRQohRhW9p2WF6dwEENpZJgJuFPRmzTzYYeLt3NK8SSBjWah58TBw-Jif_APsicA_jEW4oJnxL_yXzpIaCnQoMQj6lsNC_VmCez_bN5vun_rJVtlrKJ0S2aYkAzqZCuAkQOf1J1lGisK9cxdoa9alzAoN0cub9AFvvNw5YuQZliB5q7BKpXzvJDp_O8QiXARtBhVfrIHFV4BHg8yghLHoWJwGy-OEaIfzTSsGcoIGYNFCaUxza8hfDwo8lgohVMUOG4QJVwklgwjmP_jEulYKkJuVRI8qz6fmIvZJwO8djjMcZb1_DhL9pLt4bNuq4EUZfUcVSKfIooxv6-k6UsXb0ICSuCrzroQn00kv5kxTo-WfZ4PDGO_N5ylS44R_p_RlcEJximq4KWilKyfb90ess2nTIb26uo5udP9lXczXlhGXKF8LD6flBByZjedP8RvjTTgY6yL07qXElbc19gJDxXDZfG3FAXTofMPwTHKwRuzyfW0isyUoBhMkl2FOpInX19nMNYxWG9vNKhK9MAm_JVYojqNQsSIrKK99MFdcpW6LkgefR-BWC99IhvnvS3Vuq_jQ4jC9qSamBRYjYIQ.RK0V3dbJhOE2OCu3YgRFag/__results___files/__results___21_0.png\")
Hey everyone!
\nThis looks like it's going to be a really interesting problem!
\nI noticed that we're given the center of mass as the ground truth 'x', 'y', 'z' for the different particles and we're given a radius. Is it reasonable to assume the particles of interest will always be spherical?
\nI'm trying to generate masks for the ground truth of a starter U-Net. The radius, resolution, and center point will be sufficient if the particles are all essentially spheres, but I'm not 100% sure if this is the case. Anyone have thoughts on this?
", "messages": [ { "id": "3039408", "postDate": "11/08/2024 02:51:07", "content": "Hey everyone!
\nThis looks like it's going to be a really interesting problem!
\nI noticed that we're given the center of mass as the ground truth 'x', 'y', 'z' for the different particles and we're given a radius. Is it reasonable to assume the particles of interest will always be spherical?
\nI'm trying to generate masks for the ground truth of a starter U-Net. The radius, resolution, and center point will be sufficient if the particles are all essentially spheres, but I'm not 100% sure if this is the case. Anyone have thoughts on this?
", "rawMarkdown": "Hey everyone!\n\nThis looks like it's going to be a really interesting problem!\n\nI noticed that we're given the center of mass as the ground truth 'x', 'y', 'z' for the different particles and we're given a radius. Is it reasonable to assume the particles of interest will always be spherical?\n\nI'm trying to generate masks for the ground truth of a starter U-Net. The radius, resolution, and center point will be sufficient if the particles are all essentially spheres, but I'm not 100% sure if this is the case. Anyone have thoughts on this?", "votes": null }, { "id": "3039548", "postDate": "11/08/2024 07:11:03", "content": "Hello, if you look through the BlobDetector notebook, we’re given PDB IDs for the particles of interest, you can look them up online in 3D. Some definitely looked spherical to me, thyroglobulin does not; however, for the purposes of a starter, I’d say it’s a safe assumption to make.
", "rawMarkdown": "Hello, if you look through the BlobDetector notebook, we’re given PDB IDs for the particles of interest, you can look them up online in 3D. Some definitely looked spherical to me, thyroglobulin does [not](https://www.rcsb.org/structure/6scj); however, for the purposes of a starter, I’d say it’s a safe assumption to make.", "votes": null }, { "id": "3039738", "postDate": "11/08/2024 10:20:34", "content": "Virus particles don't look like they're always spherical. The others may be close enough though.
\nhttps://www.kaggle.com/code/davidlist/experiment-ts-6-4-visualization
\nVirus particles:
\n
Thanks David!
", "rawMarkdown": "Thanks David!", "votes": null }, { "id": "3040822", "postDate": "11/09/2024 15:57:43", "content": "Good idea to check the PDB IDs :) Thank you!
", "rawMarkdown": "Good idea to check the PDB IDs :) Thank you!", "votes": null }, { "id": "3040937", "postDate": "11/09/2024 17:32:01", "content": "your target is to predict the xyz coord, not segmentation.
\nif you treat the kaggle task like keypoint detection in human pose estimation, you just need a Gaussian/spherical point as target ground truth in the heatmap.
", "rawMarkdown": "your target is to predict the xyz coord, not segmentation.\n\nif you treat the kaggle task like keypoint detection in human pose estimation, you just need a Gaussian/spherical point as target ground truth in the heatmap.", "votes": null }, { "id": "3040943", "postDate": "11/09/2024 17:50:35", "content": "Thanks @hengck23 :) My first idea was to try and do segmentation, but I have since realized keypoint detection is likely more appropriate here. Your comment confirms to me I'm on the right track!
", "rawMarkdown": "Thanks @hengck23 :) My first idea was to try and do segmentation, but I have since realized keypoint detection is likely more appropriate here. Your comment confirms to me I'm on the right track!", "votes": null }, { "id": "3042978", "postDate": "11/12/2024 01:12:32", "content": "Actually… I'm not sure any of them are spheres. Things get weird if you view them from the side or top. Below is a ribosome which I would have guessed is the most spherical. Looks more like a cyclinder, although I suspect that's an artifact of the process of generating the tomogram, not reality.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10704200%2F5776db051bc3fb629f610a2fa016963b%2Fribosomes.png?generation=1731373948881686&alt=media\")
They aren't spheres. You'll see that apo-ferritin kinda looks like a ring. VLP is kind of like a larger ring with some extra structure. Ribosomes are kinda blobby. We've been joking that thyroglobulin kind of looks like a croissant. Beta-galactosidase has some symmetry, but also isn't a sphere.
\nThe challenge with using the particle shapes when creating masks is that you need to know the orientation of each particle. This is available with the synthetic data, but the particle orientation needs to be solved for in the experimental data.
\nIt can be useful to use a fraction of the particle of interest's radius to guarantee that the sphere is contained within the particle. In practice, we found that using spheres still works pretty well for all particle types in the competition. However, I wouldn't be surprised if a smarter way of making masks would lead to better performance.
", "rawMarkdown": "They aren't spheres. You'll see that apo-ferritin kinda looks like a ring. VLP is kind of like a larger ring with some extra structure. Ribosomes are kinda blobby. We've been joking that thyroglobulin kind of looks like a croissant. Beta-galactosidase has some symmetry, but also isn't a sphere.\n\nThe challenge with using the particle shapes when creating masks is that you need to know the orientation of each particle. This is available with the synthetic data, but the particle orientation needs to be solved for in the experimental data. \n\nIt can be useful to use a fraction of the particle of interest's radius to guarantee that the sphere is contained within the particle. In practice, we found that using spheres still works pretty well for all particle types in the competition. However, I wouldn't be surprised if a smarter way of making masks would lead to better performance.", "votes": null }, { "id": "3043997", "postDate": "11/12/2024 22:30:40", "content": "Thanks Kyle!
", "rawMarkdown": "Thanks Kyle!", "votes": null }, { "id": "3044007", "postDate": "11/12/2024 22:37:00", "content": "To add to Kyle's response: sometimes you might have multiple similar particles in a chain distributed in Z, like multiple apoferritin or multiple ribosomes. So on top of the stretching, sometimes you are looking at more than one particle actually.
", "rawMarkdown": "To add to Kyle's response: sometimes you might have multiple similar particles in a chain distributed in Z, like multiple apoferritin or multiple ribosomes. So on top of the stretching, sometimes you are looking at more than one particle actually.", "votes": null }, { "id": "3052178", "postDate": "11/22/2024 06:27:48", "content": "hello，you mean we have the radius of particles in json flie？but i did‘t found it
", "rawMarkdown": "hello，you mean we have the radius of particles in json flie？but i did‘t found it", "votes": null }, { "id": "3052186", "postDate": "11/22/2024 06:38:56", "content": "Look at the top of https://www.kaggle.com/code/kharrington/blobdetector. It has them in the json config.
", "rawMarkdown": "Look at the top of [https://www.kaggle.com/code/kharrington/blobdetector](https://www.kaggle.com/code/kharrington/blobdetector). It has them in the json config.", "votes": null }, { "id": "3052204", "postDate": "11/22/2024 07:14:38", "content": "Thank you so much for pointing that out!
", "rawMarkdown": "Thank you so much for pointing that out!", "votes": null }, { "id": "3052269", "postDate": "11/22/2024 08:22:33", "content": "If computational resources allow, you could also explore dynamic mask generation where the geometry adapts to irregular shapes based on additional input features or shape descriptors. What is your plan for handling particles that are elongated or irregular?\"
", "rawMarkdown": "If computational resources allow, you could also explore dynamic mask generation where the geometry adapts to irregular shapes based on additional input features or shape descriptors. What is your plan for handling particles that are elongated or irregular?\"", "votes": null } ], "comments": [ { "id": 3039548, "author_name": "andreizamfir", "author_url": "", "post_date": "11/08/2024 07:11:03", "content": "Hello, if you look through the BlobDetector notebook, we’re given PDB IDs for the particles of interest, you can look them up online in 3D. Some definitely looked spherical to me, thyroglobulin does not; however, for the purposes of a starter, I’d say it’s a safe assumption to make.
", "votes": null, "replies": [ { "id": 3040822, "author_name": "chemdatafarmer", "author_url": "", "post_date": "11/09/2024 15:57:43", "content": "Good idea to check the PDB IDs :) Thank you!
", "votes": null, "replies": [] } ] }, { "id": 3039738, "author_name": "davidlist", "author_url": "", "post_date": "11/08/2024 10:20:34", "content": "Virus particles don't look like they're always spherical. The others may be close enough though.
\nhttps://www.kaggle.com/code/davidlist/experiment-ts-6-4-visualization
\nVirus particles:
\n
Thanks David!
", "votes": null, "replies": [ { "id": 3042978, "author_name": "davidlist", "author_url": "", "post_date": "11/12/2024 01:12:32", "content": "Actually… I'm not sure any of them are spheres. Things get weird if you view them from the side or top. Below is a ribosome which I would have guessed is the most spherical. Looks more like a cyclinder, although I suspect that's an artifact of the process of generating the tomogram, not reality.
\n
They aren't spheres. You'll see that apo-ferritin kinda looks like a ring. VLP is kind of like a larger ring with some extra structure. Ribosomes are kinda blobby. We've been joking that thyroglobulin kind of looks like a croissant. Beta-galactosidase has some symmetry, but also isn't a sphere.
\nThe challenge with using the particle shapes when creating masks is that you need to know the orientation of each particle. This is available with the synthetic data, but the particle orientation needs to be solved for in the experimental data.
\nIt can be useful to use a fraction of the particle of interest's radius to guarantee that the sphere is contained within the particle. In practice, we found that using spheres still works pretty well for all particle types in the competition. However, I wouldn't be surprised if a smarter way of making masks would lead to better performance.
", "votes": null, "replies": [ { "id": 3043997, "author_name": "chemdatafarmer", "author_url": "", "post_date": "11/12/2024 22:30:40", "content": "Thanks Kyle!
", "votes": null, "replies": [ { "id": 3044007, "author_name": "rezaparaan", "author_url": "", "post_date": "11/12/2024 22:37:00", "content": "To add to Kyle's response: sometimes you might have multiple similar particles in a chain distributed in Z, like multiple apoferritin or multiple ribosomes. So on top of the stretching, sometimes you are looking at more than one particle actually.
", "votes": null, "replies": [] } ] } ] } ] } ] } ] }, { "id": 3040937, "author_name": "hengck23", "author_url": "", "post_date": "11/09/2024 17:32:01", "content": "your target is to predict the xyz coord, not segmentation.
\nif you treat the kaggle task like keypoint detection in human pose estimation, you just need a Gaussian/spherical point as target ground truth in the heatmap.
", "votes": null, "replies": [ { "id": 3040943, "author_name": "chemdatafarmer", "author_url": "", "post_date": "11/09/2024 17:50:35", "content": "Thanks @hengck23 :) My first idea was to try and do segmentation, but I have since realized keypoint detection is likely more appropriate here. Your comment confirms to me I'm on the right track!
", "votes": null, "replies": [] } ] }, { "id": 3052178, "author_name": "switch9527", "author_url": "", "post_date": "11/22/2024 06:27:48", "content": "hello，you mean we have the radius of particles in json flie？but i did‘t found it
", "votes": null, "replies": [ { "id": 3052186, "author_name": "davidlist", "author_url": "", "post_date": "11/22/2024 06:38:56", "content": "Look at the top of https://www.kaggle.com/code/kharrington/blobdetector. It has them in the json config.
", "votes": null, "replies": [ { "id": 3052204, "author_name": "switch9527", "author_url": "", "post_date": "11/22/2024 07:14:38", "content": "Thank you so much for pointing that out!
", "votes": null, "replies": [] } ] } ] }, { "id": 3052269, "author_name": "nancyalaswad90", "author_url": "", "post_date": "11/22/2024 08:22:33", "content": "If computational resources allow, you could also explore dynamic mask generation where the geometry adapts to irregular shapes based on additional input features or shape descriptors. What is your plan for handling particles that are elongated or irregular?\"
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3039408": "Hey everyone!\n\nThis looks like it's going to be a really interesting problem!\n\nI noticed that we're given the center of mass as the ground truth 'x', 'y', 'z' for the different particles and we're given a radius. Is it reasonable to assume the particles of interest will always be spherical?\n\nI'm trying to generate masks for the ground truth of a starter U-Net. The radius, resolution, and center point will be sufficient if the particles are all essentially spheres, but I'm not 100% sure if this is the case. Anyone have thoughts on this?", "3039548": "Hello, if you look through the BlobDetector notebook, we’re given PDB IDs for the particles of interest, you can look them up online in 3D. Some definitely looked spherical to me, thyroglobulin does [not](https://www.rcsb.org/structure/6scj); however, for the purposes of a starter, I’d say it’s a safe assumption to make.", "3039738": "Virus particles don't look like they're always spherical. The others may be close enough though.\n\n[https://www.kaggle.com/code/davidlist/experiment-ts-6-4-visualization](https://www.kaggle.com/code/davidlist/experiment-ts-6-4-visualization)\n\nVirus particles:\n", "3040821": "Thanks David!", "3040822": "Good idea to check the PDB IDs :) Thank you!", "3040937": "your target is to predict the xyz coord, not segmentation.\n\nif you treat the kaggle task like keypoint detection in human pose estimation, you just need a Gaussian/spherical point as target ground truth in the heatmap.", "3040943": "Thanks @hengck23 :) My first idea was to try and do segmentation, but I have since realized keypoint detection is likely more appropriate here. Your comment confirms to me I'm on the right track!", "3042978": "Actually... I'm not sure any of them are spheres. Things get weird if you view them from the side or top. Below is a ribosome which I would have guessed is the most spherical. Looks more like a cyclinder, although I suspect that's an artifact of the process of generating the tomogram, not reality.\n\n", "3043618": "They aren't spheres. You'll see that apo-ferritin kinda looks like a ring. VLP is kind of like a larger ring with some extra structure. Ribosomes are kinda blobby. We've been joking that thyroglobulin kind of looks like a croissant. Beta-galactosidase has some symmetry, but also isn't a sphere.\n\nThe challenge with using the particle shapes when creating masks is that you need to know the orientation of each particle. This is available with the synthetic data, but the particle orientation needs to be solved for in the experimental data. \n\nIt can be useful to use a fraction of the particle of interest's radius to guarantee that the sphere is contained within the particle. In practice, we found that using spheres still works pretty well for all particle types in the competition. However, I wouldn't be surprised if a smarter way of making masks would lead to better performance.", "3043997": "Thanks Kyle!", "3044007": "To add to Kyle's response: sometimes you might have multiple similar particles in a chain distributed in Z, like multiple apoferritin or multiple ribosomes. So on top of the stretching, sometimes you are looking at more than one particle actually.", "3052178": "hello，you mean we have the radius of particles in json flie？but i did‘t found it", "3052186": "Look at the top of [https://www.kaggle.com/code/kharrington/blobdetector](https://www.kaggle.com/code/kharrington/blobdetector). It has them in the json config.", "3052204": "Thank you so much for pointing that out!", "3052269": "If computational resources allow, you could also explore dynamic mask generation where the geometry adapts to irregular shapes based on additional input features or shape descriptors. What is your plan for handling particles that are elongated or irregular?\"" }, "source": "meta" }