First things first -- congrats to the winners! we can't wait to incorporate the work all of you guys have put into this competition into our ash to text machines.
\nI'd like to thank everyone who took time from their days to help contribute to this project. This competition was one more step towards finally being able to read the Herculaneum scrolls, without damaging them. You're all now part of a chain of research that has been ongoing for over 275 years!
\nThe predictions from models you guys have trained will help us produce real ancient greek text in the very near future.
\nThere is however a ton of work left before we can read every scroll from this ancient library. If any of you are interested in continuing to work on this , we'd be forever thankful! We're always looking for help on any of our major \"problem\" categories (more detail on these here)
\nWe're always around in our Discord and always happy to talk scrolls!
", "messages": [ { "id": 3415554, "postDate": "2026-03-01T02:11:49.653Z", "content": "First things first -- congrats to the winners! we can't wait to incorporate the work all of you guys have put into this competition into our ash to text machines.
\nI'd like to thank everyone who took time from their days to help contribute to this project. This competition was one more step towards finally being able to read the Herculaneum scrolls, without damaging them. You're all now part of a chain of research that has been ongoing for over 275 years!
\nThe predictions from models you guys have trained will help us produce real ancient greek text in the very near future.
\nThere is however a ton of work left before we can read every scroll from this ancient library. If any of you are interested in continuing to work on this , we'd be forever thankful! We're always looking for help on any of our major \"problem\" categories (more detail on these here)
\nWe're always around in our Discord and always happy to talk scrolls!
", "rawMarkdown": "First things first -- congrats to the winners! we can't wait to incorporate the work all of you guys have put into this competition into our ash to text machines. \n\nI'd like to thank everyone who took time from their days to help contribute to this project. This competition was one more step towards finally being able to read the Herculaneum scrolls, *without damaging them*. You're all now part of a chain of research that has been ongoing for over 275 years!\n\nThe predictions from models you guys have trained will help us produce real ancient greek text in the very near future. \n\nThere is however a *ton* of work left before we can read every scroll from this ancient library. If any of you are interested in continuing to work on this , we'd be forever thankful! We're always looking for help on any of our major \"problem\" categories (more detail on these [here](https://scrollprize.org/)) \n\n- ink detection\n- representations (these surface predictions being one of them)\n- meshing / unwrapping (lots of open-ended questions here) \n\nWe're always around in our [Discord](https://discord.com/invite/uTfNwwecCQ) and always happy to talk scrolls! ", "votes": 30 }, { "id": 3415734, "postDate": "2026-03-01T08:18:05.827Z", "content": "What a bumpy ride guys!
\nCongratulations to the winners, but also to everyone who participated! I am confident that most of the brilliant solutions you guys are sharing can been combined to make even better models! We are already starting looking into it, and as @seanjohnsonsp said we are eager to extract some more text, hopefully not only Greek 👀
\nI want to thank the Kaggle team, @sohier , @macruzbar and also @ryanholbrook for their constant support and enthusiasm, both before the launch and during the competition. Their advice has been very precious, especially when we had to make tough decision.
\nFinally, I think that a big shout-out should go to @seanjohnsonsp . He, with the support of other people from the annotation team, created a first version of the dataset you worked on \"by hand\". When not even a first surface model was available to create even the first pseudo labels, Sean jumped into the images and started creating some meshes which he later voxelized! He placed the dots manually, node after node. The amount of effort is really insane! When later on we had to fix some of the issues in the data, he endured my grumpiness, and worked non stop with positive attitude even during weekends to make sure that the problems were fixed, at the best of our knowledge.
\nLooking at this Vesuvius Challenge project, it's crazy how everyone is giving his 100% to make it successful. There's a mystical source of charm in these scrolls ✨
\nRegarding possible other challenges on Kaggle… we have something in mind about improving ink detection, but for the moment it's only an idea! We'll keep you posted!
", "rawMarkdown": "What a bumpy ride guys!\n\nCongratulations to the winners, but also to everyone who participated! I am confident that most of the brilliant solutions you guys are sharing can been combined to make even better models! We are already starting looking into it, and as @seanjohnsonsp said we are eager to extract some more text, hopefully not only Greek 👀\n\nI want to thank the Kaggle team, @sohier , @macruzbar and also @ryanholbrook for their constant support and enthusiasm, both before the launch and during the competition. Their advice has been very precious, especially when we had to make tough decision.\n\nFinally, I think that a big shout-out should go to @seanjohnsonsp . He, with the support of other people from the annotation team, created a first version of the dataset you worked on \"by hand\". When not even a first surface model was available to create even the first pseudo labels, Sean jumped into the images and started creating some meshes which he later voxelized! He placed the dots manually, node after node. The amount of effort is really insane! When later on we had to fix some of the issues in the data, he endured my grumpiness, and worked non stop with positive attitude even during weekends to make sure that the problems were fixed, at the best of our knowledge.\n\nLooking at this Vesuvius Challenge project, it's crazy how everyone is giving his 100% to make it successful. There's a mystical source of charm in these scrolls ✨\n\nRegarding possible other challenges on Kaggle... we have something in mind about improving ink detection, but for the moment it's only an idea! We'll keep you posted!", "votes": 12 }, { "id": 3415569, "postDate": "2026-03-01T02:27:38.817Z", "content": "Hi @seanjohnsonsp, thanks for huge effort for hosting, I would like to ask if there will be another Vesuvius Challenge on Kaggle in the future?
", "rawMarkdown": "Hi @seanjohnsonsp, thanks for huge effort for hosting, I would like to ask if there will be another Vesuvius Challenge on Kaggle in the future?", "votes": 5, "replies": [ { "id": 3415613, "postDate": "2026-03-01T03:34:46.037Z", "content": "It's something we're always considering. The challenge with hosting kaggle competitions on our data is that we must come up with a metric which cleanly maps to our goal -- reading scrolls.
\nThis recto surface prediction challenge for example was created because as of now our virtual unwrapping pipeline requires a more \"simple\" representation of the scroll, and the writing surface seems a good target for where you'd like your mesh to be, but asking a model to predict this is actually more complex than it first appears.
\nFrom the jump , what actually is the scroll surface? A new sheet of papyrus is at a macro level is comprised simply of reeds of the plant stripped and pressed together, but after many (sometimes hundreds) of years of use and handling, and at the resolution and voxel size we scan at, this descriptor no longer fits (even more so after its been blasted with pyroclastic flow) . It is not possible to discern the reeds themselves in the tomography data, and what we actually see are hundreds of different tiny structures and occlusions:
\nso when we ask the model to predict the surface, and do so with good topological properties for downstream meshing, we are really asking for something that is much harder than normal semantic segmentation, and to make matters worse it is incredibly difficult to accurately label this.
\nIf we think of something like ink detection - the actual machine learning predictions of lettered areas on the scroll - we could provide labels and use simple metrics like dice or IoU, but these fail to capture what our true target from ink detection is , which is readability. What one metric captures what an experienced papyrologist would rate as \"readable\"? What does readable even mean?
\nWe are actively working on finding answers to these questions, and when we feel comfortable with those answers enough that we can say that the results from another hosted competition would further the progress of this project it is something that we'd probably revisit.
", "rawMarkdown": "It's something we're always considering. The challenge with hosting kaggle competitions on our data is that we must come up with a metric which cleanly maps to our goal -- reading scrolls. \n\nThis recto surface prediction challenge for example was created because as of now our virtual unwrapping pipeline requires a more \"simple\" representation of the scroll, and the writing surface seems a good target for where you'd like your mesh to be, but asking a model to predict this is actually more complex than it first appears. \n\nFrom the jump , what actually *is* the scroll surface? A new sheet of papyrus is at a macro level is comprised simply of reeds of the plant stripped and pressed together, but after many (sometimes hundreds) of years of use and handling, and at the resolution and voxel size we scan at, this descriptor no longer fits (even more so after its been blasted with pyroclastic flow) . It is not possible to discern the reeds themselves in the tomography data, and what we actually see are hundreds of different tiny structures and occlusions: \n\n- vascular structures (these are the bright things we've taken to call fibers, but which are not actually fibers). these are the most dominant feature and responsible for the rectangular grid like pattern\n- hairs, bug parts, sand, dirt, even plants! \n- melted and re solidified \"crust\" of minerals and materials of all types \n- and much more \n\nso when we ask the model to predict the surface, and do so with good topological properties for downstream meshing, we are really asking for something that is much harder than normal semantic segmentation, and to make matters worse it is *incredibly* difficult to accurately label this. \n\nIf we think of something like ink detection - the actual machine learning predictions of lettered areas on the scroll - we could provide labels and use simple metrics like dice or IoU, but these fail to capture what our true target from ink detection is , which is readability. What one metric captures what an experienced papyrologist would rate as \"readable\"? What does readable even mean? \n\nWe are actively working on finding answers to these questions, and when we feel comfortable with those answers enough that we can say that the results from another hosted competition would further the progress of this project it is something that we'd probably revisit. ", "votes": 1, "replies": [ { "id": 3415622, "postDate": "2026-03-01T03:57:04.933Z", "content": "The most challenging part may be ensuring that surface predictions can ultimately be converted into revealing more ink in the scrolls. However, when hosting a competition on Kaggle, the evaluation metric must be clearly defined and implementable within the platform’s constraints, which makes metric design particularly difficult.\nTo be honest, developing an appropriate metric for this challenge is itself a research problem. I would recommend exploring a perception-based metric. For example, in the early days of generative modeling, the FID score was introduced to evaluate the quality and diversity of generated outputs. Similarly, one could train a model to assess whether a predicted surface meaningfully contributes to revealing more ink, and then use a feature-space distance measure as the evaluation metric.
", "rawMarkdown": "The most challenging part may be ensuring that surface predictions can ultimately be converted into revealing more ink in the scrolls. However, when hosting a competition on Kaggle, the evaluation metric must be clearly defined and implementable within the platform’s constraints, which makes metric design particularly difficult.\nTo be honest, developing an appropriate metric for this challenge is itself a research problem. I would recommend exploring a perception-based metric. For example, in the early days of generative modeling, the FID score was introduced to evaluate the quality and diversity of generated outputs. Similarly, one could train a model to assess whether a predicted surface meaningfully contributes to revealing more ink, and then use a feature-space distance measure as the evaluation metric." }, { "id": 3415628, "postDate": "2026-03-01T04:07:49.470Z", "content": "For this challenge thankfully the surface predictions are immediately useful ,we have unwrapping pipelines that use models trained in this same exact task already, but they are worse at it than what you guys have all been able to do. The improvements from this challenge will be great to have as our unwrapping pipelines are getting quite good but the most challenging areas in the scrolls still present road blocks, and model improvements have become difficult for us in those areas -- you all have made significant progress in those areas and that will directly translate to more unrolled scrolls.
\nTo the think metric point, yeah you're 100% right that it is a research problem in and of itself. This has been one of the more rewarding (and sometimes frustrating!) parts of working on this project. It involves so many different fields of work to go from a scroll in a museum to letters in a flattened papyrus png. Thank you for the pointer to FID and perception based metrics, i'll start digging into those tonight!
", "rawMarkdown": "For this challenge thankfully the surface predictions are immediately useful ,we have unwrapping pipelines that use models trained in this same exact task already, but they are worse at it than what you guys have all been able to do. The improvements from this challenge will be great to have as our unwrapping pipelines are getting quite good but the most challenging areas in the scrolls still present road blocks, and model improvements have become difficult for us in those areas -- you all have made significant progress in those areas and that will directly translate to more unrolled scrolls. \n\nTo the think metric point, yeah you're 100% right that it is a research problem in and of itself. This has been one of the more rewarding (and sometimes frustrating!) parts of working on this project. It involves so many different fields of work to go from a scroll in a museum to letters in a flattened papyrus png. Thank you for the pointer to FID and perception based metrics, i'll start digging into those tonight! ", "votes": 3 } ] } ] }, { "id": 3415612, "postDate": "2026-03-01T03:33:53.770Z", "content": "Thank you to the organizers for the huge effort in hosting the competition. Will there be a 3rd Vesuvius competition?
", "rawMarkdown": "Thank you to the organizers for the huge effort in hosting the competition. Will there be a 3rd Vesuvius competition?", "votes": 4, "replies": [ { "id": 3415618, "postDate": "2026-03-01T03:46:04.180Z", "content": "I wrote a rather lengthy reply to a similar question in another comment , but the long and short of it is that before we make any decision to host another one , we have to have a good metric and cleanly packaged problem which maps nicely to a kaggle comp
", "rawMarkdown": "I wrote a rather lengthy reply to a similar question in another comment , but the long and short of it is that before we make any decision to host another one , we have to have a good metric and cleanly packaged problem which maps nicely to a kaggle comp", "votes": 7 } ] }, { "id": 3415593, "postDate": "2026-03-01T02:58:38.880Z", "content": "Thank you to the organizers. This is my first time working in the field of computer vision, and I hope our team's results will be of some assistance.
", "rawMarkdown": "Thank you to the organizers. This is my first time working in the field of computer vision, and I hope our team's results will be of some assistance.", "votes": 2 }, { "id": 3415556, "postDate": "2026-03-01T02:16:35.163Z", "content": "Thank you for hosting! I learned a lot and am excited to keep up with the work you guys are doing. Though we had our ups and downs and I may have personally preferred the version where we were in the lead, I am thankful for the journey! I hope you guys were able to gets lots of improvement from our work, we tried to do something unique for you! 🙂
", "rawMarkdown": "Thank you for hosting! I learned a lot and am excited to keep up with the work you guys are doing. Though we had our ups and downs and I may have personally preferred the version where we were in the lead, I am thankful for the journey! I hope you guys were able to gets lots of improvement from our work, we tried to do something unique for you! 🙂", "replies": [ { "id": 3415559, "postDate": "2026-03-01T02:20:20.353Z", "content": "I can absolutely relate to the ups and downs! I can't tell you how many times just in a single day i've gone from feeling like i'm making no progress (and never will!) to feeling like i'll solve this problem the next day. Love seeing the creative approaches as well. There's lots of ideas that have come up that may be applicable outside of single volume recto surface predictions
", "rawMarkdown": "I can absolutely relate to the ups and downs! I can't tell you how many times just in a single day i've gone from feeling like i'm making no progress (and never will!) to feeling like i'll solve this problem the next day. Love seeing the creative approaches as well. There's lots of ideas that have come up that may be applicable outside of single volume recto surface predictions ", "votes": 2, "replies": [ { "id": 3415672, "postDate": "2026-03-01T05:20:32.863Z", "content": "Yes, I hope to help implement some of them in Vesuvius comp 3.0 ;)
", "rawMarkdown": "Yes, I hope to help implement some of them in Vesuvius comp 3.0 ;)" } ] } ] }, { "id": 3416067, "postDate": "2026-03-02T01:37:42.270Z", "content": "This competition beautifully combined machine learning and history. Proud to contribute toward decoding ancient Greek text through innovation. Looking forward to solving more challenges ahead!
", "rawMarkdown": "This competition beautifully combined machine learning and history. Proud to contribute toward decoding ancient Greek text through innovation. Looking forward to solving more challenges ahead!" }, { "id": 3415960, "postDate": "2026-03-01T17:49:26.167Z", "content": "Thank you very much to all the organizers for this amazing challenge.\nEven though I joined late, I learned a great deal from it.\nI’m very excited to follow your future work and would be happy to discuss and contribute to any challenges within it
", "rawMarkdown": "Thank you very much to all the organizers for this amazing challenge.\nEven though I joined late, I learned a great deal from it.\nI’m very excited to follow your future work and would be happy to discuss and contribute to any challenges within it" }, { "id": 3415725, "postDate": "2026-03-01T08:00:39.580Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3415734, "author_name": "Giorgio Angelotti", "author_url": "", "post_date": "2026-03-01T08:18:05.827000", "content": "What a bumpy ride guys!
\nCongratulations to the winners, but also to everyone who participated! I am confident that most of the brilliant solutions you guys are sharing can been combined to make even better models! We are already starting looking into it, and as @seanjohnsonsp said we are eager to extract some more text, hopefully not only Greek 👀
\nI want to thank the Kaggle team, @sohier , @macruzbar and also @ryanholbrook for their constant support and enthusiasm, both before the launch and during the competition. Their advice has been very precious, especially when we had to make tough decision.
\nFinally, I think that a big shout-out should go to @seanjohnsonsp . He, with the support of other people from the annotation team, created a first version of the dataset you worked on \"by hand\". When not even a first surface model was available to create even the first pseudo labels, Sean jumped into the images and started creating some meshes which he later voxelized! He placed the dots manually, node after node. The amount of effort is really insane! When later on we had to fix some of the issues in the data, he endured my grumpiness, and worked non stop with positive attitude even during weekends to make sure that the problems were fixed, at the best of our knowledge.
\nLooking at this Vesuvius Challenge project, it's crazy how everyone is giving his 100% to make it successful. There's a mystical source of charm in these scrolls ✨
\nRegarding possible other challenges on Kaggle… we have something in mind about improving ink detection, but for the moment it's only an idea! We'll keep you posted!
", "votes": 12, "replies": [] }, { "id": 3415569, "author_name": "Tom", "author_url": "", "post_date": "2026-03-01T02:27:38.817000", "content": "Hi @seanjohnsonsp, thanks for huge effort for hosting, I would like to ask if there will be another Vesuvius Challenge on Kaggle in the future?
", "votes": 5, "replies": [ { "id": 3415613, "author_name": "Sean Johnson_SP", "author_url": "", "post_date": "2026-03-01T03:34:46.037000", "content": "It's something we're always considering. The challenge with hosting kaggle competitions on our data is that we must come up with a metric which cleanly maps to our goal -- reading scrolls.
\nThis recto surface prediction challenge for example was created because as of now our virtual unwrapping pipeline requires a more \"simple\" representation of the scroll, and the writing surface seems a good target for where you'd like your mesh to be, but asking a model to predict this is actually more complex than it first appears.
\nFrom the jump , what actually is the scroll surface? A new sheet of papyrus is at a macro level is comprised simply of reeds of the plant stripped and pressed together, but after many (sometimes hundreds) of years of use and handling, and at the resolution and voxel size we scan at, this descriptor no longer fits (even more so after its been blasted with pyroclastic flow) . It is not possible to discern the reeds themselves in the tomography data, and what we actually see are hundreds of different tiny structures and occlusions:
\nso when we ask the model to predict the surface, and do so with good topological properties for downstream meshing, we are really asking for something that is much harder than normal semantic segmentation, and to make matters worse it is incredibly difficult to accurately label this.
\nIf we think of something like ink detection - the actual machine learning predictions of lettered areas on the scroll - we could provide labels and use simple metrics like dice or IoU, but these fail to capture what our true target from ink detection is , which is readability. What one metric captures what an experienced papyrologist would rate as \"readable\"? What does readable even mean?
\nWe are actively working on finding answers to these questions, and when we feel comfortable with those answers enough that we can say that the results from another hosted competition would further the progress of this project it is something that we'd probably revisit.
", "votes": 1, "replies": [ { "id": 3415622, "author_name": "Tom", "author_url": "", "post_date": "2026-03-01T03:57:04.933000", "content": "The most challenging part may be ensuring that surface predictions can ultimately be converted into revealing more ink in the scrolls. However, when hosting a competition on Kaggle, the evaluation metric must be clearly defined and implementable within the platform’s constraints, which makes metric design particularly difficult.\nTo be honest, developing an appropriate metric for this challenge is itself a research problem. I would recommend exploring a perception-based metric. For example, in the early days of generative modeling, the FID score was introduced to evaluate the quality and diversity of generated outputs. Similarly, one could train a model to assess whether a predicted surface meaningfully contributes to revealing more ink, and then use a feature-space distance measure as the evaluation metric.
", "votes": 0, "replies": [] }, { "id": 3415628, "author_name": "Sean Johnson_SP", "author_url": "", "post_date": "2026-03-01T04:07:49.470000", "content": "For this challenge thankfully the surface predictions are immediately useful ,we have unwrapping pipelines that use models trained in this same exact task already, but they are worse at it than what you guys have all been able to do. The improvements from this challenge will be great to have as our unwrapping pipelines are getting quite good but the most challenging areas in the scrolls still present road blocks, and model improvements have become difficult for us in those areas -- you all have made significant progress in those areas and that will directly translate to more unrolled scrolls.
\nTo the think metric point, yeah you're 100% right that it is a research problem in and of itself. This has been one of the more rewarding (and sometimes frustrating!) parts of working on this project. It involves so many different fields of work to go from a scroll in a museum to letters in a flattened papyrus png. Thank you for the pointer to FID and perception based metrics, i'll start digging into those tonight!
", "votes": 3, "replies": [] } ] } ] }, { "id": 3415612, "author_name": "Arunodhayan", "author_url": "", "post_date": "2026-03-01T03:33:53.770000", "content": "Thank you to the organizers for the huge effort in hosting the competition. Will there be a 3rd Vesuvius competition?
", "votes": 4, "replies": [ { "id": 3415618, "author_name": "Sean Johnson_SP", "author_url": "", "post_date": "2026-03-01T03:46:04.180000", "content": "I wrote a rather lengthy reply to a similar question in another comment , but the long and short of it is that before we make any decision to host another one , we have to have a good metric and cleanly packaged problem which maps nicely to a kaggle comp
", "votes": 7, "replies": [] } ] }, { "id": 3415593, "author_name": "GG Ayo (AyoGG)", "author_url": "", "post_date": "2026-03-01T02:58:38.880000", "content": "Thank you to the organizers. This is my first time working in the field of computer vision, and I hope our team's results will be of some assistance.
", "votes": 2, "replies": [] }, { "id": 3415556, "author_name": "Cody_Null", "author_url": "", "post_date": "2026-03-01T02:16:35.163000", "content": "Thank you for hosting! I learned a lot and am excited to keep up with the work you guys are doing. Though we had our ups and downs and I may have personally preferred the version where we were in the lead, I am thankful for the journey! I hope you guys were able to gets lots of improvement from our work, we tried to do something unique for you! 🙂
", "votes": 0, "replies": [ { "id": 3415559, "author_name": "Sean Johnson_SP", "author_url": "", "post_date": "2026-03-01T02:20:20.353000", "content": "I can absolutely relate to the ups and downs! I can't tell you how many times just in a single day i've gone from feeling like i'm making no progress (and never will!) to feeling like i'll solve this problem the next day. Love seeing the creative approaches as well. There's lots of ideas that have come up that may be applicable outside of single volume recto surface predictions
", "votes": 2, "replies": [ { "id": 3415672, "author_name": "Cody_Null", "author_url": "", "post_date": "2026-03-01T05:20:32.863000", "content": "Yes, I hope to help implement some of them in Vesuvius comp 3.0 ;)
", "votes": 0, "replies": [] } ] } ] }, { "id": 3416067, "author_name": "MahnoorKhan.", "author_url": "", "post_date": "2026-03-02T01:37:42.270000", "content": "This competition beautifully combined machine learning and history. Proud to contribute toward decoding ancient Greek text through innovation. Looking forward to solving more challenges ahead!
", "votes": 0, "replies": [] }, { "id": 3415960, "author_name": "Eng Adam Al mohammedi", "author_url": "", "post_date": "2026-03-01T17:49:26.167000", "content": "Thank you very much to all the organizers for this amazing challenge.\nEven though I joined late, I learned a great deal from it.\nI’m very excited to follow your future work and would be happy to discuss and contribute to any challenges within it
", "votes": 0, "replies": [] }, { "id": 3415725, "author_name": "", "author_url": "", "post_date": "2026-03-01T08:00:39.580000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3415554": "First things first -- congrats to the winners! we can't wait to incorporate the work all of you guys have put into this competition into our ash to text machines. \n\nI'd like to thank everyone who took time from their days to help contribute to this project. This competition was one more step towards finally being able to read the Herculaneum scrolls, *without damaging them*. You're all now part of a chain of research that has been ongoing for over 275 years!\n\nThe predictions from models you guys have trained will help us produce real ancient greek text in the very near future. \n\nThere is however a *ton* of work left before we can read every scroll from this ancient library. If any of you are interested in continuing to work on this , we'd be forever thankful! We're always looking for help on any of our major \"problem\" categories (more detail on these [here](https://scrollprize.org/)) \n\n- ink detection\n- representations (these surface predictions being one of them)\n- meshing / unwrapping (lots of open-ended questions here) \n\nWe're always around in our [Discord](https://discord.com/invite/uTfNwwecCQ) and always happy to talk scrolls! ", "3415734": "What a bumpy ride guys!\n\nCongratulations to the winners, but also to everyone who participated! I am confident that most of the brilliant solutions you guys are sharing can been combined to make even better models! We are already starting looking into it, and as @seanjohnsonsp said we are eager to extract some more text, hopefully not only Greek 👀\n\nI want to thank the Kaggle team, @sohier , @macruzbar and also @ryanholbrook for their constant support and enthusiasm, both before the launch and during the competition. Their advice has been very precious, especially when we had to make tough decision.\n\nFinally, I think that a big shout-out should go to @seanjohnsonsp . He, with the support of other people from the annotation team, created a first version of the dataset you worked on \"by hand\". When not even a first surface model was available to create even the first pseudo labels, Sean jumped into the images and started creating some meshes which he later voxelized! He placed the dots manually, node after node. The amount of effort is really insane! When later on we had to fix some of the issues in the data, he endured my grumpiness, and worked non stop with positive attitude even during weekends to make sure that the problems were fixed, at the best of our knowledge.\n\nLooking at this Vesuvius Challenge project, it's crazy how everyone is giving his 100% to make it successful. There's a mystical source of charm in these scrolls ✨\n\nRegarding possible other challenges on Kaggle... we have something in mind about improving ink detection, but for the moment it's only an idea! We'll keep you posted!", "3415569": "Hi @seanjohnsonsp, thanks for huge effort for hosting, I would like to ask if there will be another Vesuvius Challenge on Kaggle in the future?", "3415612": "Thank you to the organizers for the huge effort in hosting the competition. Will there be a 3rd Vesuvius competition?", "3415593": "Thank you to the organizers. This is my first time working in the field of computer vision, and I hope our team's results will be of some assistance.", "3415556": "Thank you for hosting! I learned a lot and am excited to keep up with the work you guys are doing. Though we had our ups and downs and I may have personally preferred the version where we were in the lead, I am thankful for the journey! I hope you guys were able to gets lots of improvement from our work, we tried to do something unique for you! 🙂", "3416067": "This competition beautifully combined machine learning and history. Proud to contribute toward decoding ancient Greek text through innovation. Looking forward to solving more challenges ahead!", "3415960": "Thank you very much to all the organizers for this amazing challenge.\nEven though I joined late, I learned a great deal from it.\nI’m very excited to follow your future work and would be happy to discuss and contribute to any challenges within it", "3415725": "" } }