a localization network can be helpful to crop only relevant part of the images, this would be much better than center-cropping, which tends to loss information
", "messages": [ { "id": "116859", "postDate": "04/26/2016 10:15:58", "content": "a localization network can be helpful to crop only relevant part of the images, this would be much better than center-cropping, which tends to loss information
", "rawMarkdown": "a localization network can be helpful to crop only relevant part of the images, this would be much better than center-cropping, which tends to loss information", "votes": null }, { "id": "116866", "postDate": "04/26/2016 11:04:35", "content": "I was thinking about the same thing! How can we do this? Do we need to use OpenCV ? Any starting points?
", "rawMarkdown": "I was thinking about the same thing! How can we do this? Do we need to use OpenCV ? Any starting points?", "votes": null }, { "id": "117092", "postDate": "04/27/2016 11:20:57", "content": "Its based on a pretrained CNN, http://pjreddie.com/darknet/yolo/,
", "rawMarkdown": "Its based on a pretrained CNN, http://pjreddie.com/darknet/yolo/,", "votes": null }, { "id": "117107", "postDate": "04/27/2016 12:42:45", "content": "Did you have success with this method ?\nWith this logic, wouldn't segmentation be better ?
", "rawMarkdown": "Did you have success with this method ?\r\nWith this logic, wouldn't segmentation be better ?", "votes": null }, { "id": "117109", "postDate": "04/27/2016 12:47:47", "content": "[quote=Jonathan Chung;117107]
\n\nDid you have success with this method ?\nWith this logic, wouldn't segmentation be better ?
\n\n[/quote]\nNot yet, currently, tried with finetuning and feature extraction from fine tuned model. What do you mean by segmentation? Which part to segment ?
", "rawMarkdown": "[quote=Jonathan Chung;117107]\r\n\r\nDid you have success with this method ?\r\nWith this logic, wouldn't segmentation be better ?\r\n\r\n[/quote]\r\nNot yet, currently, tried with finetuning and feature extraction from fine tuned model. What do you mean by segmentation? Which part to segment ?", "votes": null }, { "id": "117120", "postDate": "04/27/2016 13:28:48", "content": "I suppose we can say segmentation is the location of "human" pixels. Although I haven't tried it, I believe it's reasonably accurate.
\n\n[quote=SecondPlan;117109]
\n\n[quote=Jonathan Chung;117107]
\n\nDid you have success with this method ?\nWith this logic, wouldn't segmentation be better ?
\n\n[/quote]\nNot yet, currently, tried with finetuning and feature extraction from fine tuned model. What do you mean by segmentation? Which part to segment ?
\n\n[/quote]
", "rawMarkdown": "I suppose we can say segmentation is the location of \"human\" pixels. Although I haven't tried it, I believe it's reasonably accurate.\r\n \r\n[quote=SecondPlan;117109]\r\n\r\n[quote=Jonathan Chung;117107]\r\n\r\nDid you have success with this method ?\r\nWith this logic, wouldn't segmentation be better ?\r\n\r\n[/quote]\r\nNot yet, currently, tried with finetuning and feature extraction from fine tuned model. What do you mean by segmentation? Which part to segment ?\r\n\r\n[/quote]", "votes": null }, { "id": "117480", "postDate": "04/29/2016 09:23:24", "content": "well, what's the difference between the 'human' pixels and my posted figures, which are exactly the box of humans
", "rawMarkdown": "well, what's the difference between the 'human' pixels and my posted figures, which are exactly the box of humans", "votes": null }, { "id": "117581", "postDate": "04/29/2016 18:24:27", "content": "An example of segmentation is in here 1. As you can see it's not really a "bounding box".
", "rawMarkdown": "An example of segmentation is in here [1]. As you can see it's not really a \"bounding box\".\r\n\r\n [1]: http://www.robots.ox.ac.uk/~szheng/crfasrnndemo", "votes": null }, { "id": "117609", "postDate": "04/29/2016 20:29:15", "content": "[quote=Jonathan Chung;117581]
\n\nAn example of segmentation is in here 1. As you can see it's not really a "bounding box".
\n\n[/quote]\nHey, that's interesting, I would invest some time looking into that. Thanks, man
", "rawMarkdown": "[quote=Jonathan Chung;117581]\r\n\r\nAn example of segmentation is in here [1]. As you can see it's not really a \"bounding box\".\r\n\r\n [1]: http://www.robots.ox.ac.uk/~szheng/crfasrnndemo\r\n\r\n[/quote]\r\nHey, that's interesting, I would invest some time looking into that. Thanks, man", "votes": null }, { "id": "117660", "postDate": "04/30/2016 02:25:02", "content": "[quote=Jonathan Chung;117581]
\n\nAn example of segmentation is in here 1. As you can see it's not really a "bounding box".
\n\n[/quote]
\n\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?
", "rawMarkdown": "[quote=Jonathan Chung;117581]\r\n\r\nAn example of segmentation is in here [1]. As you can see it's not really a \"bounding box\".\r\n\r\n [1]: http://www.robots.ox.ac.uk/~szheng/crfasrnndemo\r\n\r\n[/quote]\r\n\r\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?", "votes": null }, { "id": "117706", "postDate": "04/30/2016 13:12:00", "content": "[quote=Bojan Tunguz;117660]
\n\n[quote=Jonathan Chung;117581]
\n\nAn example of segmentation is in here 1. As you can see it's not really a "bounding box".
\n\n[/quote]
\n\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?
\n\n[/quote]
\n\nI suppose that's a possibility. I believe the segmentation should assign a label towards "human pixels" so one can remove other unnecessary information (e.g. the chair).\nThat being said, I haven't really explored it in detail as the net takes lots of memory.
", "rawMarkdown": "[quote=Bojan Tunguz;117660]\r\n\r\n[quote=Jonathan Chung;117581]\r\n\r\nAn example of segmentation is in here [1]. As you can see it's not really a \"bounding box\".\r\n\r\n [1]: http://www.robots.ox.ac.uk/~szheng/crfasrnndemo\r\n\r\n[/quote]\r\n\r\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?\r\n\r\n[/quote]\r\n\r\nI suppose that's a possibility. I believe the segmentation should assign a label towards \"human pixels\" so one can remove other unnecessary information (e.g. the chair).\r\nThat being said, I haven't really explored it in detail as the net takes lots of memory.", "votes": null }, { "id": "117736", "postDate": "04/30/2016 16:11:50", "content": "[quote=Jonathan Chung;117706]
\n\n[quote=Bojan Tunguz;117660]
\n\n[quote=Jonathan Chung;117581]
\n\nAn example of segmentation is in here 1. As you can see it's not really a "bounding box".
\n\n[/quote]
\n\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?
\n\n[/quote]
\n\nI suppose that's a possibility. I believe the segmentation should assign a label towards "human pixels" so one can remove other unnecessary information (e.g. the chair).\nThat being said, I haven't really explored it in detail as the net takes lots of memory.
\n\n[/quote]
\n\nThese are some sample plots of the segmented image vs original one and there are some classes that looks like not that easy to be separated
", "rawMarkdown": "[quote=Jonathan Chung;117706]\r\n\r\n[quote=Bojan Tunguz;117660]\r\n\r\n[quote=Jonathan Chung;117581]\r\n\r\nAn example of segmentation is in here [1]. As you can see it's not really a \"bounding box\".\r\n\r\n [1]: http://www.robots.ox.ac.uk/~szheng/crfasrnndemo\r\n\r\n[/quote]\r\n\r\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?\r\n\r\n[/quote]\r\n\r\nI suppose that's a possibility. I believe the segmentation should assign a label towards \"human pixels\" so one can remove other unnecessary information (e.g. the chair).\r\nThat being said, I haven't really explored it in detail as the net takes lots of memory.\r\n\r\n\r\n[/quote]\r\n\r\n\r\nThese are some sample plots of the segmented image vs original one and there are some classes that looks like not that easy to be separated", "votes": null }, { "id": "118901", "postDate": "05/06/2016 00:48:16", "content": "[quote=SecondPlan;117092]
\n\nIts based on a pretrained CNN, http://pjreddie.com/darknet/yolo/,
\n\n[/quote]
\n\nHave you been able to get the coordinates of the cropping rectangles? The scripts they provide you with there only seem to allow for saving of the .png images with drawn rectangles.
", "rawMarkdown": "[quote=SecondPlan;117092]\r\n\r\nIts based on a pretrained CNN, http://pjreddie.com/darknet/yolo/,\r\n\r\n[/quote]\r\n\r\nHave you been able to get the coordinates of the cropping rectangles? The scripts they provide you with there only seem to allow for saving of the .png images with drawn rectangles.", "votes": null }, { "id": "119101", "postDate": "05/07/2016 07:45:04", "content": "you need to change the c code and recompile it
", "rawMarkdown": "you need to change the c code and recompile it", "votes": null }, { "id": "124905", "postDate": "06/23/2016 09:39:46", "content": "", "rawMarkdown": "", "votes": null }, { "id": "125114", "postDate": "06/26/2016 12:17:45", "content": "Hi SecondPlan,
\n\nThank you for sharing interesting idea.
\nRemoving unnecessary information may help us.
Please forget about segmentation, there are better ways to crop images. I am attaching a link of cropped images. \n100 images for 10classes
", "rawMarkdown": "Please forget about segmentation, there are better ways to crop images. I am attaching a link of cropped images. \r\n[100 images for 10classes][1]\r\n\r\n\r\n [1]: https://www.dropbox.com/s/qyvfd0c8ruaw01v/samples.zip?dl=0", "votes": null }, { "id": "125576", "postDate": "06/30/2016 13:58:36", "content": "@SecondPlan: So you are saying cropping based on Yolo output doesn't help?
\n\nI tried a less sophisticated approach for cropping (based on classifying the car), and it didn't have significant impact on the performance.
", "rawMarkdown": "SecondPlan: So you are saying cropping based on Yolo output doesn't help?\r\n\r\n I tried a less sophisticated approach for cropping (based on classifying the car), and it didn't have significant impact on the performance.", "votes": null }, { "id": "125858", "postDate": "07/03/2016 15:01:08", "content": "@Alexander Bauer\nIf you are thinking of localization, etc, you can consider my ideas here:\nhttps://www.kaggle.com/c/state-farm-distracted-driver-detection/forums/t/21994/heat-map-of-cnn-output
\n\n[quote=Alexander Bauer;125576]
\n\n@SecondPlan: So you are saying cropping based on Yolo output doesn't help?
\n\nI tried a less sophisticated approach for cropping (based on classifying the car), and it didn't have significant impact on the performance.
\n\n[/quote]
", "rawMarkdown": "Alexander Bauer\r\nIf you are thinking of localization, etc, you can consider my ideas here:\r\nhttps://www.kaggle.com/c/state-farm-distracted-driver-detection/forums/t/21994/heat-map-of-cnn-output\r\n\r\n[quote=Alexander Bauer;125576]\r\n\r\n@SecondPlan: So you are saying cropping based on Yolo output doesn't help?\r\n\r\n I tried a less sophisticated approach for cropping (based on classifying the car), and it didn't have significant impact on the performance.\r\n\r\n[/quote]", "votes": null } ], "comments": [ { "id": 116866, "author_name": "abhijayvuyyuru", "author_url": "", "post_date": "04/26/2016 11:04:35", "content": "I was thinking about the same thing! How can we do this? Do we need to use OpenCV ? Any starting points?
", "votes": null, "replies": [] }, { "id": 117092, "author_name": "usixuz", "author_url": "", "post_date": "04/27/2016 11:20:57", "content": "Its based on a pretrained CNN, http://pjreddie.com/darknet/yolo/,
", "votes": null, "replies": [] }, { "id": 117107, "author_name": "jonomon", "author_url": "", "post_date": "04/27/2016 12:42:45", "content": "Did you have success with this method ?\nWith this logic, wouldn't segmentation be better ?
", "votes": null, "replies": [] }, { "id": 117109, "author_name": "usixuz", "author_url": "", "post_date": "04/27/2016 12:47:47", "content": "[quote=Jonathan Chung;117107]
\n\nDid you have success with this method ?\nWith this logic, wouldn't segmentation be better ?
\n\n[/quote]\nNot yet, currently, tried with finetuning and feature extraction from fine tuned model. What do you mean by segmentation? Which part to segment ?
", "votes": null, "replies": [] }, { "id": 117120, "author_name": "jonomon", "author_url": "", "post_date": "04/27/2016 13:28:48", "content": "I suppose we can say segmentation is the location of "human" pixels. Although I haven't tried it, I believe it's reasonably accurate.
\n\n[quote=SecondPlan;117109]
\n\n[quote=Jonathan Chung;117107]
\n\nDid you have success with this method ?\nWith this logic, wouldn't segmentation be better ?
\n\n[/quote]\nNot yet, currently, tried with finetuning and feature extraction from fine tuned model. What do you mean by segmentation? Which part to segment ?
\n\n[/quote]
", "votes": null, "replies": [] }, { "id": 117480, "author_name": "usixuz", "author_url": "", "post_date": "04/29/2016 09:23:24", "content": "well, what's the difference between the 'human' pixels and my posted figures, which are exactly the box of humans
", "votes": null, "replies": [] }, { "id": 117581, "author_name": "jonomon", "author_url": "", "post_date": "04/29/2016 18:24:27", "content": "An example of segmentation is in here 1. As you can see it's not really a "bounding box".
", "votes": null, "replies": [] }, { "id": 117609, "author_name": "usixuz", "author_url": "", "post_date": "04/29/2016 20:29:15", "content": "[quote=Jonathan Chung;117581]
\n\nAn example of segmentation is in here 1. As you can see it's not really a "bounding box".
\n\n[/quote]\nHey, that's interesting, I would invest some time looking into that. Thanks, man
", "votes": null, "replies": [] }, { "id": 117660, "author_name": "tunguz", "author_url": "", "post_date": "04/30/2016 02:25:02", "content": "[quote=Jonathan Chung;117581]
\n\nAn example of segmentation is in here 1. As you can see it's not really a "bounding box".
\n\n[/quote]
\n\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?
", "votes": null, "replies": [] }, { "id": 117706, "author_name": "jonomon", "author_url": "", "post_date": "04/30/2016 13:12:00", "content": "[quote=Bojan Tunguz;117660]
\n\n[quote=Jonathan Chung;117581]
\n\nAn example of segmentation is in here 1. As you can see it's not really a "bounding box".
\n\n[/quote]
\n\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?
\n\n[/quote]
\n\nI suppose that's a possibility. I believe the segmentation should assign a label towards "human pixels" so one can remove other unnecessary information (e.g. the chair).\nThat being said, I haven't really explored it in detail as the net takes lots of memory.
", "votes": null, "replies": [] }, { "id": 117736, "author_name": "usixuz", "author_url": "", "post_date": "04/30/2016 16:11:50", "content": "[quote=Jonathan Chung;117706]
\n\n[quote=Bojan Tunguz;117660]
\n\n[quote=Jonathan Chung;117581]
\n\nAn example of segmentation is in here 1. As you can see it's not really a "bounding box".
\n\n[/quote]
\n\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?
\n\n[/quote]
\n\nI suppose that's a possibility. I believe the segmentation should assign a label towards "human pixels" so one can remove other unnecessary information (e.g. the chair).\nThat being said, I haven't really explored it in detail as the net takes lots of memory.
\n\n[/quote]
\n\nThese are some sample plots of the segmented image vs original one and there are some classes that looks like not that easy to be separated
", "votes": null, "replies": [] }, { "id": 118901, "author_name": "tunguz", "author_url": "", "post_date": "05/06/2016 00:48:16", "content": "[quote=SecondPlan;117092]
\n\nIts based on a pretrained CNN, http://pjreddie.com/darknet/yolo/,
\n\n[/quote]
\n\nHave you been able to get the coordinates of the cropping rectangles? The scripts they provide you with there only seem to allow for saving of the .png images with drawn rectangles.
", "votes": null, "replies": [] }, { "id": 119101, "author_name": "usixuz", "author_url": "", "post_date": "05/07/2016 07:45:04", "content": "you need to change the c code and recompile it
", "votes": null, "replies": [] }, { "id": 124905, "author_name": "vdpappu", "author_url": "", "post_date": "06/23/2016 09:39:46", "content": "", "votes": null, "replies": [] }, { "id": 125114, "author_name": "toshik", "author_url": "", "post_date": "06/26/2016 12:17:45", "content": "Hi SecondPlan,
\n\nThank you for sharing interesting idea.
\nRemoving unnecessary information may help us.
Please forget about segmentation, there are better ways to crop images. I am attaching a link of cropped images. \n100 images for 10classes
", "votes": null, "replies": [] }, { "id": 125576, "author_name": "allexius", "author_url": "", "post_date": "06/30/2016 13:58:36", "content": "@SecondPlan: So you are saying cropping based on Yolo output doesn't help?
\n\nI tried a less sophisticated approach for cropping (based on classifying the car), and it didn't have significant impact on the performance.
", "votes": null, "replies": [] }, { "id": 125858, "author_name": "hengck23", "author_url": "", "post_date": "07/03/2016 15:01:08", "content": "@Alexander Bauer\nIf you are thinking of localization, etc, you can consider my ideas here:\nhttps://www.kaggle.com/c/state-farm-distracted-driver-detection/forums/t/21994/heat-map-of-cnn-output
\n\n[quote=Alexander Bauer;125576]
\n\n@SecondPlan: So you are saying cropping based on Yolo output doesn't help?
\n\nI tried a less sophisticated approach for cropping (based on classifying the car), and it didn't have significant impact on the performance.
\n\n[/quote]
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "116859": "a localization network can be helpful to crop only relevant part of the images, this would be much better than center-cropping, which tends to loss information", "116866": "I was thinking about the same thing! How can we do this? Do we need to use OpenCV ? Any starting points?", "117092": "Its based on a pretrained CNN, http://pjreddie.com/darknet/yolo/,", "117107": "Did you have success with this method ?\r\nWith this logic, wouldn't segmentation be better ?", "117109": "[quote=Jonathan Chung;117107]\r\n\r\nDid you have success with this method ?\r\nWith this logic, wouldn't segmentation be better ?\r\n\r\n[/quote]\r\nNot yet, currently, tried with finetuning and feature extraction from fine tuned model. What do you mean by segmentation? Which part to segment ?", "117120": "I suppose we can say segmentation is the location of \"human\" pixels. Although I haven't tried it, I believe it's reasonably accurate.\r\n \r\n[quote=SecondPlan;117109]\r\n\r\n[quote=Jonathan Chung;117107]\r\n\r\nDid you have success with this method ?\r\nWith this logic, wouldn't segmentation be better ?\r\n\r\n[/quote]\r\nNot yet, currently, tried with finetuning and feature extraction from fine tuned model. What do you mean by segmentation? Which part to segment ?\r\n\r\n[/quote]", "117480": "well, what's the difference between the 'human' pixels and my posted figures, which are exactly the box of humans", "117581": "An example of segmentation is in here [1]. As you can see it's not really a \"bounding box\".\r\n\r\n [1]: http://www.robots.ox.ac.uk/~szheng/crfasrnndemo", "117609": "[quote=Jonathan Chung;117581]\r\n\r\nAn example of segmentation is in here [1]. As you can see it's not really a \"bounding box\".\r\n\r\n [1]: http://www.robots.ox.ac.uk/~szheng/crfasrnndemo\r\n\r\n[/quote]\r\nHey, that's interesting, I would invest some time looking into that. Thanks, man", "117660": "[quote=Jonathan Chung;117581]\r\n\r\nAn example of segmentation is in here [1]. As you can see it's not really a \"bounding box\".\r\n\r\n [1]: http://www.robots.ox.ac.uk/~szheng/crfasrnndemo\r\n\r\n[/quote]\r\n\r\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?", "117706": "[quote=Bojan Tunguz;117660]\r\n\r\n[quote=Jonathan Chung;117581]\r\n\r\nAn example of segmentation is in here [1]. As you can see it's not really a \"bounding box\".\r\n\r\n [1]: http://www.robots.ox.ac.uk/~szheng/crfasrnndemo\r\n\r\n[/quote]\r\n\r\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?\r\n\r\n[/quote]\r\n\r\nI suppose that's a possibility. I believe the segmentation should assign a label towards \"human pixels\" so one can remove other unnecessary information (e.g. the chair).\r\nThat being said, I haven't really explored it in detail as the net takes lots of memory.", "117736": "[quote=Jonathan Chung;117706]\r\n\r\n[quote=Bojan Tunguz;117660]\r\n\r\n[quote=Jonathan Chung;117581]\r\n\r\nAn example of segmentation is in here [1]. As you can see it's not really a \"bounding box\".\r\n\r\n [1]: http://www.robots.ox.ac.uk/~szheng/crfasrnndemo\r\n\r\n[/quote]\r\n\r\nA quick question: after running segmentation would you then just assign a value of 0 to all non-segment pixels, and then train your NN on thus modified images?\r\n\r\n[/quote]\r\n\r\nI suppose that's a possibility. I believe the segmentation should assign a label towards \"human pixels\" so one can remove other unnecessary information (e.g. the chair).\r\nThat being said, I haven't really explored it in detail as the net takes lots of memory.\r\n\r\n\r\n[/quote]\r\n\r\n\r\nThese are some sample plots of the segmented image vs original one and there are some classes that looks like not that easy to be separated", "118901": "[quote=SecondPlan;117092]\r\n\r\nIts based on a pretrained CNN, http://pjreddie.com/darknet/yolo/,\r\n\r\n[/quote]\r\n\r\nHave you been able to get the coordinates of the cropping rectangles? The scripts they provide you with there only seem to allow for saving of the .png images with drawn rectangles.", "119101": "you need to change the c code and recompile it", "124905": "", "125114": "Hi SecondPlan,\r\n\r\nThank you for sharing interesting idea.