In couple weeks I will give a talk about our solution at a Meetup in San Francisco.
\n\nhttps://www.meetup.com/San-Francisco-Kagglers/events/239113801/
\n\nI already wrote a blog post in Russian that describes our approach, problems that we were facing and how did we overcome them.
\n\nIn couple weeks I will give a talk about our solution at a Meetup in San Francisco.
\n\nhttps://www.meetup.com/San-Francisco-Kagglers/events/239113801/
\n\nI already wrote a blog post in Russian that describes our approach, problems that we were facing and how did we overcome them.
\n\nExcellent. I shall be there, if my schedule cooperates.
\n\nAny other DSTL'ers in the Bay Area? We should go out afterwards for drinks, and complain to each other about submission hell.
", "rawMarkdown": "Excellent. I shall be there, if my schedule cooperates. \n\nAny other DSTL'ers in the Bay Area? We should go out afterwards for drinks, and complain to each other about submission hell.", "votes": null }, { "id": "174262", "postDate": "04/10/2017 20:49:40", "content": "Thanks for writing up your solution. I look forward to the English version. I'll give Google Translate a try until then.
\n\nI'll see you at the meetup.
", "rawMarkdown": "Thanks for writing up your solution. I look forward to the English version. I'll give Google Translate a try until then.\n\n1. You mentioned [Tiramisu][1] in the article. Can you say more about any attempts to apply it to this problem?\n2. What do you think about [Mask R-CNN][2] for this problem?\n\nI'll see you at the meetup.\n\n\n [1]: https://arxiv.org/abs/1611.09326\n [2]: https://arxiv.org/abs/1703.06870", "votes": null }, { "id": "174263", "postDate": "04/10/2017 20:51:29", "content": "Hey, I know your teammate, mfab. Are you taking Part II of fast.ai's deep learning course?
", "rawMarkdown": "Hey, I know your teammate, mfab. Are you taking Part II of [fast.ai's deep learning course][1]?\n\n\n [1]: http://course.fast.ai/", "votes": null }, { "id": "174264", "postDate": "04/10/2017 20:52:19", "content": "It would be great to see you there.
\n\nTo my knowledge, there we many participants from Russia, and they already had drinks in Moscow, Saint Petersburg and other cities with similar discussions. :)
\n\nKaggle is not really popular in the Bay Area, I mean, nearly every Data Scientist heard about it, but I would say only one tenth of one percent tried to do an actual submission. => I am not expecting more than 2-5 people who actually tried to work on this problem at the meetup.
", "rawMarkdown": "It would be great to see you there.\n\nTo my knowledge, there we many participants from Russia, and they already had drinks in Moscow, Saint Petersburg and other cities with similar discussions. :)\n\nKaggle is not really popular in the Bay Area, I mean, nearly every Data Scientist heard about it, but I would say only one tenth of one percent tried to do an actual submission. => I am not expecting more than 2-5 people who actually tried to work on this problem at the meetup.", "votes": null }, { "id": "174267", "postDate": "04/10/2017 21:01:09", "content": "Russian version is just a story that describes a path to the final solution.
\n\nI am planning to finish blog post for a Kaggle forum this week, and it will be much richer with technical details.
\n\nWe tried to use Tiramisu, but it showed similar to Unet performance, but due to its complexity it was converging much slower. @alno had some success with a simplified version of it, but I did not ask him about details.
Mask R-CNN looks interesting and I am planning to implement it, at least to check how does it work on the ImageNet dataset that is ultra clean and nice. Maybe when I will get comfortable with it I will try to apply it to this problem.
mfab is taking the fast.ai course. I'm following along vicariously. But it was Jeremy Howard's discussion of kaggle that sucked me into giving DSTL a go.
", "rawMarkdown": "mfab is taking the fast.ai course. I'm following along vicariously. But it was Jeremy Howard's discussion of kaggle that sucked me into giving DSTL a go.", "votes": null }, { "id": "174480", "postDate": "04/11/2017 16:18:53", "content": "Brother, share the code for others to learn as well. Thanks.
", "rawMarkdown": "Brother, share the code for others to learn as well. Thanks.", "votes": null }, { "id": "174485", "postDate": "04/11/2017 16:31:19", "content": "I am willing to share the code, but we can not get a clear answer from organizers if it is allowed by the winner's license.
", "rawMarkdown": "I am willing to share the code, but we can not get a clear answer from organizers if it is allowed by the winner's license.", "votes": null }, { "id": "174486", "postDate": "04/11/2017 16:39:38", "content": "\n\n\nRussian version is just a story that describes a path to the final solution.
\n
It was quite an enjoyable story:) I would read more reports(and learn more) if people wrote the path to the solution instead of describing just the final result.
\n\nOne thing that I also wondered about is how often are the solutions we find on arxiv, claiming state of the art simply overfit architectures and methods to the benchmark set. It is often difficult to find the little hacks they had to use buried in the technical reports.
", "rawMarkdown": "> Russian version is just a story that describes a path to the final solution. \n\nIt was quite an enjoyable story:) I would read more reports(and learn more) if people wrote the path to the solution instead of describing just the final result. \n\nOne thing that I also wondered about is how often are the solutions we find on arxiv, claiming state of the art simply overfit architectures and methods to the benchmark set. It is often difficult to find the little hacks they had to use buried in the technical reports.", "votes": null }, { "id": "174501", "postDate": "04/11/2017 18:05:08", "content": "ok, but if you get positive reply from them then do share code for us. I personally think that they should not have any problem with the sharing but anyways. i will wait. Thanks.
", "rawMarkdown": "ok, but if you get positive reply from them then do share code for us. I personally think that they should not have any problem with the sharing but anyways. i will wait. Thanks.", "votes": null }, { "id": "175901", "postDate": "04/18/2017 03:13:57", "content": "That will be great!
", "rawMarkdown": "That will be great!", "votes": null }, { "id": "179452", "postDate": "05/01/2017 15:07:01", "content": "looking forwards for English version of blog post, thanks a lot.
", "rawMarkdown": "looking forwards for English version of blog post, thanks a lot.", "votes": null }, { "id": "179457", "postDate": "05/01/2017 15:19:33", "content": "We already submitted it to kaggle. I believe it is a question of a few days.
", "rawMarkdown": "We already submitted it to kaggle. I believe it is a question of a few days.", "votes": null }, { "id": "179733", "postDate": "05/02/2017 17:34:10", "content": "Video: https://www.youtube.com/watch?v=lCcLjPOCof0
\n\nPresentation: http://slides.com/vladimiriglovikov/using-deep-learning-for-satellite-imagery-feature-detection-challenge/fullscreen
", "rawMarkdown": "Video: https://www.youtube.com/watch?v=lCcLjPOCof0\n\nPresentation: http://slides.com/vladimiriglovikov/using-deep-learning-for-satellite-imagery-feature-detection-challenge/fullscreen", "votes": null }, { "id": "179835", "postDate": "05/03/2017 00:54:36", "content": "Neat, I'd love to be there!
", "rawMarkdown": "Neat, I'd love to be there!", "votes": null }, { "id": "179852", "postDate": "05/03/2017 03:04:00", "content": "Talk will be repeated in the H2O headquarters on May 24th.
\n\nhttps://www.meetup.com/Silicon-Valley-Big-Data-Science/events/239683067/
", "rawMarkdown": "Talk will be repeated in the H2O headquarters on May 24th.\n\nhttps://www.meetup.com/Silicon-Valley-Big-Data-Science/events/239683067/", "votes": null }, { "id": "180788", "postDate": "05/07/2017 08:08:54", "content": "@Vladimir, Thank you for the meetup video. Will you share the code please.
", "rawMarkdown": "Vladimir, Thank you for the meetup video. Will you share the code please.", "votes": null }, { "id": "262423", "postDate": "12/26/2017 14:46:16", "content": "", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 174260, "author_name": "threeplusone", "author_url": "", "post_date": "04/10/2017 20:40:22", "content": "Excellent. I shall be there, if my schedule cooperates.
\n\nAny other DSTL'ers in the Bay Area? We should go out afterwards for drinks, and complain to each other about submission hell.
", "votes": null, "replies": [ { "id": 174263, "author_name": "kleinsmith", "author_url": "", "post_date": "04/10/2017 20:51:29", "content": "Hey, I know your teammate, mfab. Are you taking Part II of fast.ai's deep learning course?
", "votes": null, "replies": [] }, { "id": 174264, "author_name": "iglovikov", "author_url": "", "post_date": "04/10/2017 20:52:19", "content": "It would be great to see you there.
\n\nTo my knowledge, there we many participants from Russia, and they already had drinks in Moscow, Saint Petersburg and other cities with similar discussions. :)
\n\nKaggle is not really popular in the Bay Area, I mean, nearly every Data Scientist heard about it, but I would say only one tenth of one percent tried to do an actual submission. => I am not expecting more than 2-5 people who actually tried to work on this problem at the meetup.
", "votes": null, "replies": [] }, { "id": 174270, "author_name": "threeplusone", "author_url": "", "post_date": "04/10/2017 21:08:00", "content": "mfab is taking the fast.ai course. I'm following along vicariously. But it was Jeremy Howard's discussion of kaggle that sucked me into giving DSTL a go.
", "votes": null, "replies": [] } ] }, { "id": 174262, "author_name": "kleinsmith", "author_url": "", "post_date": "04/10/2017 20:49:40", "content": "Thanks for writing up your solution. I look forward to the English version. I'll give Google Translate a try until then.
\n\nI'll see you at the meetup.
", "votes": null, "replies": [ { "id": 174267, "author_name": "iglovikov", "author_url": "", "post_date": "04/10/2017 21:01:09", "content": "Russian version is just a story that describes a path to the final solution.
\n\nI am planning to finish blog post for a Kaggle forum this week, and it will be much richer with technical details.
\n\nWe tried to use Tiramisu, but it showed similar to Unet performance, but due to its complexity it was converging much slower. @alno had some success with a simplified version of it, but I did not ask him about details.
Mask R-CNN looks interesting and I am planning to implement it, at least to check how does it work on the ImageNet dataset that is ultra clean and nice. Maybe when I will get comfortable with it I will try to apply it to this problem.
\n\n\nRussian version is just a story that describes a path to the final solution.
\n
It was quite an enjoyable story:) I would read more reports(and learn more) if people wrote the path to the solution instead of describing just the final result.
\n\nOne thing that I also wondered about is how often are the solutions we find on arxiv, claiming state of the art simply overfit architectures and methods to the benchmark set. It is often difficult to find the little hacks they had to use buried in the technical reports.
", "votes": null, "replies": [] } ] }, { "id": 174480, "author_name": "samihaq", "author_url": "", "post_date": "04/11/2017 16:18:53", "content": "Brother, share the code for others to learn as well. Thanks.
", "votes": null, "replies": [ { "id": 174485, "author_name": "iglovikov", "author_url": "", "post_date": "04/11/2017 16:31:19", "content": "I am willing to share the code, but we can not get a clear answer from organizers if it is allowed by the winner's license.
", "votes": null, "replies": [] }, { "id": 174501, "author_name": "samihaq", "author_url": "", "post_date": "04/11/2017 18:05:08", "content": "ok, but if you get positive reply from them then do share code for us. I personally think that they should not have any problem with the sharing but anyways. i will wait. Thanks.
", "votes": null, "replies": [] }, { "id": 175901, "author_name": "binoruv", "author_url": "", "post_date": "04/18/2017 03:13:57", "content": "That will be great!
", "votes": null, "replies": [] } ] }, { "id": 179452, "author_name": "craymond", "author_url": "", "post_date": "05/01/2017 15:07:01", "content": "looking forwards for English version of blog post, thanks a lot.
", "votes": null, "replies": [ { "id": 179457, "author_name": "iglovikov", "author_url": "", "post_date": "05/01/2017 15:19:33", "content": "We already submitted it to kaggle. I believe it is a question of a few days.
", "votes": null, "replies": [] } ] }, { "id": 179733, "author_name": "iglovikov", "author_url": "", "post_date": "05/02/2017 17:34:10", "content": "Video: https://www.youtube.com/watch?v=lCcLjPOCof0
\n\nPresentation: http://slides.com/vladimiriglovikov/using-deep-learning-for-satellite-imagery-feature-detection-challenge/fullscreen
", "votes": null, "replies": [ { "id": 179835, "author_name": "arnowaczynski", "author_url": "", "post_date": "05/03/2017 00:54:36", "content": "Neat, I'd love to be there!
", "votes": null, "replies": [] }, { "id": 262423, "author_name": "zeliek", "author_url": "", "post_date": "12/26/2017 14:46:16", "content": "", "votes": null, "replies": [] } ] }, { "id": 179852, "author_name": "iglovikov", "author_url": "", "post_date": "05/03/2017 03:04:00", "content": "Talk will be repeated in the H2O headquarters on May 24th.
\n\nhttps://www.meetup.com/Silicon-Valley-Big-Data-Science/events/239683067/
", "votes": null, "replies": [] }, { "id": 180788, "author_name": "quest13", "author_url": "", "post_date": "05/07/2017 08:08:54", "content": "@Vladimir, Thank you for the meetup video. Will you share the code please.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "174251": "In couple weeks I will give a talk about our solution at a Meetup in San Francisco.\n\nhttps://www.meetup.com/San-Francisco-Kagglers/events/239113801/\n\nI already wrote a [blog post in Russian][1] that describes our approach, problems that we were facing and how did we overcome them. \n\n\n* UPD blog post: http://blog.kaggle.com/2017/05/09/dstl-satellite-imagery-competition-3rd-place-winners-interview-vladimir-sergey/\n* UPD preprint: https://arxiv.org/abs/1706.06169\n\n [1]: https://habrahabr.ru/company/ods/blog/325096/", "174260": "Excellent. I shall be there, if my schedule cooperates. \n\nAny other DSTL'ers in the Bay Area? We should go out afterwards for drinks, and complain to each other about submission hell.", "174262": "Thanks for writing up your solution. I look forward to the English version. I'll give Google Translate a try until then.\n\n1. You mentioned [Tiramisu][1] in the article. Can you say more about any attempts to apply it to this problem?\n2. What do you think about [Mask R-CNN][2] for this problem?\n\nI'll see you at the meetup.\n\n\n [1]: https://arxiv.org/abs/1611.09326\n [2]: https://arxiv.org/abs/1703.06870", "174263": "Hey, I know your teammate, mfab. Are you taking Part II of [fast.ai's deep learning course][1]?\n\n\n [1]: http://course.fast.ai/", "174264": "It would be great to see you there.\n\nTo my knowledge, there we many participants from Russia, and they already had drinks in Moscow, Saint Petersburg and other cities with similar discussions. :)\n\nKaggle is not really popular in the Bay Area, I mean, nearly every Data Scientist heard about it, but I would say only one tenth of one percent tried to do an actual submission. => I am not expecting more than 2-5 people who actually tried to work on this problem at the meetup.", "174267": "Russian version is just a story that describes a path to the final solution. \n\nI am planning to finish blog post for a Kaggle forum this week, and it will be much richer with technical details.\n\n1. We tried to use Tiramisu, but it showed similar to Unet performance, but due to its complexity it was converging much slower. @alno had some success with a simplified version of it, but I did not ask him about details.\n\n2. Mask R-CNN looks interesting and I am planning to implement it, at least to check how does it work on the ImageNet dataset that is ultra clean and nice. Maybe when I will get comfortable with it I will try to apply it to this problem.", "174270": "mfab is taking the fast.ai course. I'm following along vicariously. But it was Jeremy Howard's discussion of kaggle that sucked me into giving DSTL a go.", "174480": "Brother, share the code for others to learn as well. Thanks.", "174485": "I am willing to share the code, but we can not get a clear answer from organizers if it is allowed by the winner's license.", "174486": "> Russian version is just a story that describes a path to the final solution. \n\nIt was quite an enjoyable story:) I would read more reports(and learn more) if people wrote the path to the solution instead of describing just the final result. \n\nOne thing that I also wondered about is how often are the solutions we find on arxiv, claiming state of the art simply overfit architectures and methods to the benchmark set. It is often difficult to find the little hacks they had to use buried in the technical reports.", "174501": "ok, but if you get positive reply from them then do share code for us. I personally think that they should not have any problem with the sharing but anyways. i will wait. Thanks.", "175901": "That will be great!", "179452": "looking forwards for English version of blog post, thanks a lot.", "179457": "We already submitted it to kaggle. I believe it is a question of a few days.", "179733": "Video: https://www.youtube.com/watch?v=lCcLjPOCof0\n\nPresentation: http://slides.com/vladimiriglovikov/using-deep-learning-for-satellite-imagery-feature-detection-challenge/fullscreen", "179835": "Neat, I'd love to be there!", "179852": "Talk will be repeated in the H2O headquarters on May 24th.\n\nhttps://www.meetup.com/Silicon-Valley-Big-Data-Science/events/239683067/", "180788": "Vladimir, Thank you for the meetup video. Will you share the code please.", "262423": "" }, "source": "meta" }