There hasn't been a ton of notebooks featuring ship localization, and most \"detection\" scripts only performs basic clssification.
\n\nWhat are the motives behind this practice? Or is speed the key point of considertion?
", "messages": [ { "id": "395860", "postDate": "09/29/2018 13:46:29", "content": "There hasn't been a ton of notebooks featuring ship localization, and most \"detection\" scripts only performs basic clssification.
\n\nWhat are the motives behind this practice? Or is speed the key point of considertion?
", "rawMarkdown": "There hasn't been a ton of notebooks featuring ship localization, and most \"detection\" scripts only performs basic clssification.\n\nWhat are the motives behind this practice? Or is speed the key point of considertion?", "votes": null }, { "id": "396567", "postDate": "10/01/2018 02:28:16", "content": "I expect other approaches will be posted after the competition is over.
", "rawMarkdown": "I expect other approaches will be posted after the competition is over.", "votes": null }, { "id": "398375", "postDate": "10/04/2018 03:09:31", "content": "I have been trying localization with a sliding window approach. One serious problem that I encounter is ships which are mooring alongside each other. I don't know how to catch more than one ship pr window with localization.
", "rawMarkdown": "I have been trying localization with a sliding window approach. One serious problem that I encounter is ships which are mooring alongside each other. I don't know how to catch more than one ship pr window with localization.", "votes": null }, { "id": "398390", "postDate": "10/04/2018 04:06:38", "content": "You can find a helpful method from this paper https://arxiv.org/pdf/1711.09405.pdf considering rotating bounding boxes.
", "rawMarkdown": "You can find a helpful method from this paper https://arxiv.org/pdf/1711.09405.pdf considering rotating bounding boxes.", "votes": null }, { "id": "398591", "postDate": "10/04/2018 09:58:17", "content": "Nice! Thank you very much. I did not know of the SSD procedure. It could be an important stepping stone.\nHowever, I wonder how fine a grid of predetermined positions I would need? The center distance between some of the small boat are only 8-10 pixel. As I understand it, that would mean 768/10 = 77 * 77 = 5929 predetermined positions. Is that realistic?
", "rawMarkdown": "Nice! Thank you very much. I did not know of the SSD procedure. It could be an important stepping stone.\nHowever, I wonder how fine a grid of predetermined positions I would need? The center distance between some of the small boat are only 8-10 pixel. As I understand it, that would mean 768/10 = 77 * 77 = 5929 predetermined positions. Is that realistic?", "votes": null }, { "id": "398606", "postDate": "10/04/2018 10:22:00", "content": "yes
", "rawMarkdown": "yes", "votes": null }, { "id": "398777", "postDate": "10/04/2018 15:13:52", "content": "You should also remember that you will need to create several anchor boxes with different aspect ratio and angle...
", "rawMarkdown": "You should also remember that you will need to create several anchor boxes with different aspect ratio and angle...", "votes": null }, { "id": "399143", "postDate": "10/05/2018 09:29:12", "content": "Yes, there will be additional factors on the number of parameters. I hope to find a way around the rotational parameters, which seems to be a pretty naive implementation. Still the other parameters remain. Perhaps Ming, could you explain why you are so optimistic?
", "rawMarkdown": "Yes, there will be additional factors on the number of parameters. I hope to find a way around the rotational parameters, which seems to be a pretty naive implementation. Still the other parameters remain. Perhaps Ming, could you explain why you are so optimistic?", "votes": null }, { "id": "399339", "postDate": "10/05/2018 16:21:11", "content": "Im not being optimistic or anything. its basically how one step detectors work. From what Im picking up youre probably doubting if using that many bounding boxes is a feasible idea.
\n\nMany one shot detectors generate thousands of bounding boxes or more (depending on the size of your input images and the anchor boxes you are using)
\n\nYou might want to learn how SSDs work its not possible to achieve this scale of things otherwise.
\n\nFun fact, the retinanet model for small object detection uses over 140000 boxes on a typical 420*560 images on coco (though they upscale the images to atleast 800 a side)
", "rawMarkdown": "Im not being optimistic or anything. its basically how one step detectors work. From what Im picking up youre probably doubting if using that many bounding boxes is a feasible idea.\n\nMany one shot detectors generate thousands of bounding boxes or more (depending on the size of your input images and the anchor boxes you are using)\n\nYou might want to learn how SSDs work its not possible to achieve this scale of things otherwise.\n\nFun fact, the retinanet model for small object detection uses over 140000 boxes on a typical 420*560 images on coco (though they upscale the images to atleast 800 a side)", "votes": null }, { "id": "399341", "postDate": "10/05/2018 16:21:53", "content": "No mistake its over a tenth of a million
", "rawMarkdown": "No mistake its over a tenth of a million", "votes": null }, { "id": "399737", "postDate": "10/06/2018 15:58:43", "content": "Ok, thank you for your informative answer.
\n\nI have re-read the paper and I now notice that Lei Liu et al. separates the original image into overlapping 300x300 pixels images each with 38x38 predetermined points giving an 8 pixels 'resolution'. Not unlike what we need here. Their total number of prior boxes including different aspect ratios and angles ends at 43320.
\n\nTheir receptive field is only 108 pixels. The largest Airbus ships are 300-400 pixels so in that respect we have to change something in the network, compared to their version.
", "rawMarkdown": "Ok, thank you for your informative answer. \n\nI have re-read the paper and I now notice that Lei Liu et al. separates the original image into overlapping 300x300 pixels images each with 38x38 predetermined points giving an 8 pixels 'resolution'. Not unlike what we need here. Their total number of prior boxes including different aspect ratios and angles ends at 43320.\n\nTheir receptive field is only 108 pixels. The largest Airbus ships are 300-400 pixels so in that respect we have to change something in the network, compared to their version.", "votes": null } ], "comments": [ { "id": 396567, "author_name": "iafoss", "author_url": "", "post_date": "10/01/2018 02:28:16", "content": "I expect other approaches will be posted after the competition is over.
", "votes": null, "replies": [] }, { "id": 398375, "author_name": "petersorensen360", "author_url": "", "post_date": "10/04/2018 03:09:31", "content": "I have been trying localization with a sliding window approach. One serious problem that I encounter is ships which are mooring alongside each other. I don't know how to catch more than one ship pr window with localization.
", "votes": null, "replies": [ { "id": 398390, "author_name": "iafoss", "author_url": "", "post_date": "10/04/2018 04:06:38", "content": "You can find a helpful method from this paper https://arxiv.org/pdf/1711.09405.pdf considering rotating bounding boxes.
", "votes": null, "replies": [] }, { "id": 398591, "author_name": "petersorensen360", "author_url": "", "post_date": "10/04/2018 09:58:17", "content": "Nice! Thank you very much. I did not know of the SSD procedure. It could be an important stepping stone.\nHowever, I wonder how fine a grid of predetermined positions I would need? The center distance between some of the small boat are only 8-10 pixel. As I understand it, that would mean 768/10 = 77 * 77 = 5929 predetermined positions. Is that realistic?
", "votes": null, "replies": [] }, { "id": 398606, "author_name": "mingruimingrui", "author_url": "", "post_date": "10/04/2018 10:22:00", "content": "yes
", "votes": null, "replies": [] }, { "id": 398777, "author_name": "iafoss", "author_url": "", "post_date": "10/04/2018 15:13:52", "content": "You should also remember that you will need to create several anchor boxes with different aspect ratio and angle...
", "votes": null, "replies": [] }, { "id": 399143, "author_name": "petersorensen360", "author_url": "", "post_date": "10/05/2018 09:29:12", "content": "Yes, there will be additional factors on the number of parameters. I hope to find a way around the rotational parameters, which seems to be a pretty naive implementation. Still the other parameters remain. Perhaps Ming, could you explain why you are so optimistic?
", "votes": null, "replies": [] }, { "id": 399339, "author_name": "mingruimingrui", "author_url": "", "post_date": "10/05/2018 16:21:11", "content": "Im not being optimistic or anything. its basically how one step detectors work. From what Im picking up youre probably doubting if using that many bounding boxes is a feasible idea.
\n\nMany one shot detectors generate thousands of bounding boxes or more (depending on the size of your input images and the anchor boxes you are using)
\n\nYou might want to learn how SSDs work its not possible to achieve this scale of things otherwise.
\n\nFun fact, the retinanet model for small object detection uses over 140000 boxes on a typical 420*560 images on coco (though they upscale the images to atleast 800 a side)
", "votes": null, "replies": [] }, { "id": 399341, "author_name": "mingruimingrui", "author_url": "", "post_date": "10/05/2018 16:21:53", "content": "No mistake its over a tenth of a million
", "votes": null, "replies": [] }, { "id": 399737, "author_name": "petersorensen360", "author_url": "", "post_date": "10/06/2018 15:58:43", "content": "Ok, thank you for your informative answer.
\n\nI have re-read the paper and I now notice that Lei Liu et al. separates the original image into overlapping 300x300 pixels images each with 38x38 predetermined points giving an 8 pixels 'resolution'. Not unlike what we need here. Their total number of prior boxes including different aspect ratios and angles ends at 43320.
\n\nTheir receptive field is only 108 pixels. The largest Airbus ships are 300-400 pixels so in that respect we have to change something in the network, compared to their version.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "395860": "There hasn't been a ton of notebooks featuring ship localization, and most \"detection\" scripts only performs basic clssification.\n\nWhat are the motives behind this practice? Or is speed the key point of considertion?", "396567": "I expect other approaches will be posted after the competition is over.", "398375": "I have been trying localization with a sliding window approach. One serious problem that I encounter is ships which are mooring alongside each other. I don't know how to catch more than one ship pr window with localization.", "398390": "You can find a helpful method from this paper https://arxiv.org/pdf/1711.09405.pdf considering rotating bounding boxes.", "398591": "Nice! Thank you very much. I did not know of the SSD procedure. It could be an important stepping stone.\nHowever, I wonder how fine a grid of predetermined positions I would need? The center distance between some of the small boat are only 8-10 pixel. As I understand it, that would mean 768/10 = 77 * 77 = 5929 predetermined positions. Is that realistic?", "398606": "yes", "398777": "You should also remember that you will need to create several anchor boxes with different aspect ratio and angle...", "399143": "Yes, there will be additional factors on the number of parameters. I hope to find a way around the rotational parameters, which seems to be a pretty naive implementation. Still the other parameters remain. Perhaps Ming, could you explain why you are so optimistic?", "399339": "Im not being optimistic or anything. its basically how one step detectors work. From what Im picking up youre probably doubting if using that many bounding boxes is a feasible idea.\n\nMany one shot detectors generate thousands of bounding boxes or more (depending on the size of your input images and the anchor boxes you are using)\n\nYou might want to learn how SSDs work its not possible to achieve this scale of things otherwise.\n\nFun fact, the retinanet model for small object detection uses over 140000 boxes on a typical 420*560 images on coco (though they upscale the images to atleast 800 a side)", "399341": "No mistake its over a tenth of a million", "399737": "Ok, thank you for your informative answer. \n\nI have re-read the paper and I now notice that Lei Liu et al. separates the original image into overlapping 300x300 pixels images each with 38x38 predetermined points giving an 8 pixels 'resolution'. Not unlike what we need here. Their total number of prior boxes including different aspect ratios and angles ends at 43320.\n\nTheir receptive field is only 108 pixels. The largest Airbus ships are 300-400 pixels so in that respect we have to change something in the network, compared to their version." }, "source": "meta" }