![\"diagram\"](\"https://raw.githubusercontent.com/toshi-k/kaggle-airbus-ship-detection-challenge/master/img/diagram.png\")
I implemented Oriented SSD for this competition. Though LB score is not so good (226th), this model have interesting characteristics. For example, it ensure that predicted shape must be \"rectangle\".
\n\nCode: https://github.com/toshi-k/kaggle-airbus-ship-detection-challenge\n
I implemented Oriented SSD for this competition. Though LB score is not so good (226th), this model have interesting characteristics. For example, it ensure that predicted shape must be \"rectangle\".
\n\nCode: https://github.com/toshi-k/kaggle-airbus-ship-detection-challenge\n
Thanks for sharing !
", "rawMarkdown": "Thanks for sharing !", "votes": null }, { "id": "424739", "postDate": "11/20/2018 15:47:25", "content": "Thank you for sharing! \nI was effectively looking for kagglers who implemented the brand new \"Rotatable Bounding Box\" detector! \nNice work. \nJeff.
", "rawMarkdown": "Thank you for sharing! \nI was effectively looking for kagglers who implemented the brand new \"Rotatable Bounding Box\" detector! \nNice work. \nJeff.", "votes": null }, { "id": "425263", "postDate": "11/21/2018 11:07:43", "content": "Hi Jeff,
\n\nThank you for hosting interesting competition. \nI thought this competition was chance to try Rotatable Bounding Box, because most target masks were rectangle.\nWhen you can't use segmentation by some reason, or want to estimate coordinates of ship directly, please use my solution.
", "rawMarkdown": "Hi Jeff,\n\nThank you for hosting interesting competition. \nI thought this competition was chance to try Rotatable Bounding Box, because most target masks were rectangle.\nWhen you can't use segmentation by some reason, or want to estimate coordinates of ship directly, please use my solution.", "votes": null }, { "id": "425506", "postDate": "11/21/2018 17:33:52", "content": "Thank you for sharing, I wanted to try this method too, but somehow got too busy and discouraged by the leak.
\n\nDo you have any ideas, why this model does not perform well in this competition? I'm asking because the post-processing our team used is also based on generation of boxes based on Unet prediction followed by their rendering as masks. This technique works really well in public LB (and visual analysis also confirms it), but lowers the score in private LB quite a lot, so our best prediction ~0.851 private LB score, which we did not select because of really low ~0.740 public LB score, was just an ensemble of Unet models without any postprocessing.
", "rawMarkdown": "Thank you for sharing, I wanted to try this method too, but somehow got too busy and discouraged by the leak. \n\nDo you have any ideas, why this model does not perform well in this competition? I'm asking because the post-processing our team used is also based on generation of boxes based on Unet prediction followed by their rendering as masks. This technique works really well in public LB (and visual analysis also confirms it), but lowers the score in private LB quite a lot, so our best prediction ~0.851 private LB score, which we did not select because of really low ~0.740 public LB score, was just an ensemble of Unet models without any postprocessing.", "votes": null }, { "id": "425662", "postDate": "11/21/2018 23:27:42", "content": "Hi Iafoss,
\n\nI wonder same things with you. The masks form U-Net have rounded corners and seem like banana 🍌. One of the reason may be evaluation metric. In this competition, the metric (average F2 score) evaluate pixel-wised unions, so the shape of masks are not so important.
", "rawMarkdown": "Hi Iafoss,\n\nI wonder same things with you. The masks form U-Net have rounded corners and seem like banana 🍌. One of the reason may be evaluation metric. In this competition, the metric (average F2 score) evaluate pixel-wised unions, so the shape of masks are not so important.", "votes": null }, { "id": "427686", "postDate": "11/26/2018 00:33:14", "content": "Congrats @toshi_k, and thanks for sharing.
", "rawMarkdown": "Congrats @toshi_k, and thanks for sharing.", "votes": null }, { "id": "1046811", "postDate": "10/12/2020 03:03:05", "content": "Thanks for sharing !😊
", "rawMarkdown": "Thanks for sharing !😊", "votes": null } ], "comments": [ { "id": 1046811, "author_name": "heiheihei2017", "author_url": "", "post_date": "10/12/2020 03:03:05", "content": "Thanks for sharing !😊
", "votes": null, "replies": [] }, { "id": 424122, "author_name": "tkuanlun", "author_url": "", "post_date": "11/19/2018 15:56:49", "content": "Thanks for sharing !
", "votes": null, "replies": [] }, { "id": 424739, "author_name": "jeffaudi", "author_url": "", "post_date": "11/20/2018 15:47:25", "content": "Thank you for sharing! \nI was effectively looking for kagglers who implemented the brand new \"Rotatable Bounding Box\" detector! \nNice work. \nJeff.
", "votes": null, "replies": [ { "id": 425263, "author_name": "toshik", "author_url": "", "post_date": "11/21/2018 11:07:43", "content": "Hi Jeff,
\n\nThank you for hosting interesting competition. \nI thought this competition was chance to try Rotatable Bounding Box, because most target masks were rectangle.\nWhen you can't use segmentation by some reason, or want to estimate coordinates of ship directly, please use my solution.
", "votes": null, "replies": [] } ] }, { "id": 425506, "author_name": "iafoss", "author_url": "", "post_date": "11/21/2018 17:33:52", "content": "Thank you for sharing, I wanted to try this method too, but somehow got too busy and discouraged by the leak.
\n\nDo you have any ideas, why this model does not perform well in this competition? I'm asking because the post-processing our team used is also based on generation of boxes based on Unet prediction followed by their rendering as masks. This technique works really well in public LB (and visual analysis also confirms it), but lowers the score in private LB quite a lot, so our best prediction ~0.851 private LB score, which we did not select because of really low ~0.740 public LB score, was just an ensemble of Unet models without any postprocessing.
", "votes": null, "replies": [ { "id": 425662, "author_name": "toshik", "author_url": "", "post_date": "11/21/2018 23:27:42", "content": "Hi Iafoss,
\n\nI wonder same things with you. The masks form U-Net have rounded corners and seem like banana 🍌. One of the reason may be evaluation metric. In this competition, the metric (average F2 score) evaluate pixel-wised unions, so the shape of masks are not so important.
", "votes": null, "replies": [] } ] }, { "id": 427686, "author_name": "sheriytm", "author_url": "", "post_date": "11/26/2018 00:33:14", "content": "Congrats @toshi_k, and thanks for sharing.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "423167": "I implemented Oriented SSD for this competition. Though LB score is not so good (226th), this model have interesting characteristics. For example, it ensure that predicted shape must be \"rectangle\".\n\nCode: https://github.com/toshi-k/kaggle-airbus-ship-detection-challenge\n![diagram][1]\n\n\n [1]: https://raw.githubusercontent.com/toshi-k/kaggle-airbus-ship-detection-challenge/master/img/diagram.png", "424122": "Thanks for sharing !", "424739": "Thank you for sharing! \nI was effectively looking for kagglers who implemented the brand new \"Rotatable Bounding Box\" detector! \nNice work. \nJeff.", "425263": "Hi Jeff,\n\nThank you for hosting interesting competition. \nI thought this competition was chance to try Rotatable Bounding Box, because most target masks were rectangle.\nWhen you can't use segmentation by some reason, or want to estimate coordinates of ship directly, please use my solution.", "425506": "Thank you for sharing, I wanted to try this method too, but somehow got too busy and discouraged by the leak. \n\nDo you have any ideas, why this model does not perform well in this competition? I'm asking because the post-processing our team used is also based on generation of boxes based on Unet prediction followed by their rendering as masks. This technique works really well in public LB (and visual analysis also confirms it), but lowers the score in private LB quite a lot, so our best prediction ~0.851 private LB score, which we did not select because of really low ~0.740 public LB score, was just an ensemble of Unet models without any postprocessing.", "425662": "Hi Iafoss,\n\nI wonder same things with you. The masks form U-Net have rounded corners and seem like banana 🍌. One of the reason may be evaluation metric. In this competition, the metric (average F2 score) evaluate pixel-wised unions, so the shape of masks are not so important.", "427686": "Congrats @toshi_k, and thanks for sharing.", "1046811": "Thanks for sharing !😊" }, "source": "meta" }