So I've been wondering if anybody's using techniques other than LBP model described by organizers. Personally I immediately rushed for a convolutional network, and surprisingly, ended up with nice enough looking results... although, of course, not perfect. The false positive cases have already been briefly discussed before, I guess - those situations where a head is obscured and some other body part (like a tail) is surrounded by foam and makes a distinct white-gray pattern.
\n\nAnd then there's an idea - maybe extracting and processing local head/background patches just doesn't give you enough information. Maybe we have to consider the spatial stucture of the body first - and direct our attention at one of it's narrow ends. The body should be relatively easy to detect; and then it's only a matter of determining its orientation and distinguishing a head from a tail...
", "messages": [ { "id": "93511", "postDate": "09/26/2015 22:35:04", "content": "So I've been wondering if anybody's using techniques other than LBP model described by organizers. Personally I immediately rushed for a convolutional network, and surprisingly, ended up with nice enough looking results... although, of course, not perfect. The false positive cases have already been briefly discussed before, I guess - those situations where a head is obscured and some other body part (like a tail) is surrounded by foam and makes a distinct white-gray pattern.
\n\nAnd then there's an idea - maybe extracting and processing local head/background patches just doesn't give you enough information. Maybe we have to consider the spatial stucture of the body first - and direct our attention at one of it's narrow ends. The body should be relatively easy to detect; and then it's only a matter of determining its orientation and distinguishing a head from a tail...
", "rawMarkdown": "So I've been wondering if anybody's using techniques other than LBP model described by organizers. Personally I immediately rushed for a convolutional network, and surprisingly, ended up with nice enough looking results... although, of course, not perfect. The false positive cases have already been briefly discussed before, I guess - those situations where a head is obscured and some other body part (like a tail) is surrounded by foam and makes a distinct white-gray pattern. \r\n\r\nAnd then there's an idea - maybe extracting and processing local head/background patches just doesn't give you enough information. Maybe we have to consider the spatial stucture of the body first - and direct our attention at one of it's narrow ends. The body should be relatively easy to detect; and then it's only a matter of determining its orientation and distinguishing a head from a tail...", "votes": null }, { "id": "93546", "postDate": "09/27/2015 19:14:05", "content": "[quote=Artem Khurshudov;93511]
\n\nPersonally I immediately rushed for a convolutional network, and surprisingly, ended up with nice enough looking results... although, of course, not perfect.
\n\n[/quote]
\n\nIf you don't mind saying, are you localizing using a convnet by regressing a bounding box or are you using a convnet as a classifier and doing some kind of sliding window approach? The results are pretty good! I am going to try the regression approach.
", "rawMarkdown": "[quote=Artem Khurshudov;93511]\r\n\r\nPersonally I immediately rushed for a convolutional network, and surprisingly, ended up with nice enough looking results... although, of course, not perfect. \r\n\r\n[/quote]\r\n\r\nIf you don't mind saying, are you localizing using a convnet by regressing a bounding box or are you using a convnet as a classifier and doing some kind of sliding window approach? The results are pretty good! I am going to try the regression approach.", "votes": null }, { "id": "93548", "postDate": "09/27/2015 20:51:54", "content": "I use the sliding window approach, and actually I don't like it much - it still produces a lot of false detections, so I had to set a threshold for network's output value. And by "regressing a bounding box" do you mean something like Yann LeCun's OverFeat framework, or do you plan to make a network that directly outputs bounding box parameters? I thought about the latter myself, but not quite sure that'll work...
", "rawMarkdown": "I use the sliding window approach, and actually I don't like it much - it still produces a lot of false detections, so I had to set a threshold for network's output value. And by \"regressing a bounding box\" do you mean something like Yann LeCun's [OverFeat][1] framework, or do you plan to make a network that directly outputs bounding box parameters? I thought about the latter myself, but not quite sure that'll work...\r\n\r\n\r\n [1]: http://arxiv.org/abs/1312.6229", "votes": null }, { "id": "93550", "postDate": "09/27/2015 20:55:28", "content": "I mean more like the R-CNN approach. So getting candidate detections from some other method, e.g. cascades, and then simultaneously mitigating false alarms and regressing the bounding box using a CNN
", "rawMarkdown": "I mean more like the R-CNN approach. So getting candidate detections from some other method, e.g. cascades, and then simultaneously mitigating false alarms and regressing the bounding box using a CNN", "votes": null }, { "id": "93551", "postDate": "09/27/2015 20:59:23", "content": "Huh. Never heard of it before, so thanks, I guess I have a paper to read tonight. :-)
", "rawMarkdown": "Huh. Never heard of it before, so thanks, I guess I have a paper to read tonight. :-)", "votes": null }, { "id": "97367", "postDate": "10/26/2015 14:22:17", "content": "So, a little update on my approach, if anyone's interested: instead of using sliding windows/OverFeat/RCNN I focused on a mask regression technique described by Szegedy et al. Basically you make a binary black and white mask where everything inside the object's bounding box is white and everything else is black. Then you train the network to predict a mask from raw pixel image (obviously, you need to prepare a training set of masks). I spent almost 3 weeks trying to make make it work, and the key to success seems to be augmenting your training set 10x more and adding lots of dropout, and then even a fairly shallow convnet does the trick.
\n\nThen another idea I dwelled on for a while was to detect not the head specifically but the whole whale body instead, because I thought it will be much easier since heads are sometimes hardly distinguishable from the water-and-foam pattern. And here are some results: "mask1-2-3" displays mask prediction accuracy and how a bounded box is fitted to the predicted mask again (nothing fancy, some linear algebra and PCA), and "predicted_bbox-1-2-3" demonstrates the final result so far, separating whales from background.
\n\n...as a side note, anybody who isn't in the team yet and likes the mask regression approach, is interested to team up? I'm a PhD student in machine learning and computer vision, who uses python/Theano/Lasagne/Keras, and this is my first Kaggle competition, so not much experience so far.
", "rawMarkdown": "So, a little update on my approach, if anyone's interested: instead of using sliding windows/OverFeat/RCNN I focused on a mask regression technique described by [Szegedy et al][1]. Basically you make a binary black and white mask where everything inside the object's bounding box is white and everything else is black. Then you train the network to predict a mask from raw pixel image (obviously, you need to prepare a training set of masks). I spent almost 3 weeks trying to make make it work, and the key to success seems to be augmenting your training set 10x more and adding lots of dropout, and then even a fairly shallow convnet does the trick. \r\n\r\nThen another idea I dwelled on for a while was to detect not the head specifically but the whole whale body instead, because I thought it will be much easier since heads are sometimes hardly distinguishable from the water-and-foam pattern. And here are some results: \"mask1-2-3\" displays mask prediction accuracy and how a bounded box is fitted to the predicted mask again (nothing fancy, some linear algebra and PCA), and \"predicted_bbox-1-2-3\" demonstrates the final result so far, separating whales from background.\r\n\r\n...as a side note, anybody who isn't in the team yet and likes the mask regression approach, is interested to team up? I'm a PhD student in machine learning and computer vision, who uses python/Theano/Lasagne/Keras, and this is my first Kaggle competition, so not much experience so far.\r\n\r\n [1]: http://papers.nips.cc/paper/5207-deep-neural-networks-for-object-detection.pdf", "votes": null }, { "id": "97386", "postDate": "10/26/2015 20:02:32", "content": "Artem,
\n\nThe results you got seem interesting. I'm interested to team up if you are open.
\n\nBasically, I'm trying to build a ConvNet using Matlab (I know this is a hard task, but not impossible). I have decent computing power to do some test runs on my end. On the side, I'm also trying to build a similar technique to isolate the whales but your method seems to produce better results.
\n\nWe can take this conversation private if you have any further questions. I hope to hear from you soon.
\n\nThanks
", "rawMarkdown": "Artem, \r\n\r\nThe results you got seem interesting. I'm interested to team up if you are open.\r\n\r\nBasically, I'm trying to build a ConvNet using Matlab (I know this is a hard task, but not impossible). I have decent computing power to do some test runs on my end. On the side, I'm also trying to build a similar technique to isolate the whales but your method seems to produce better results.\r\n\r\nWe can take this conversation private if you have any further questions. I hope to hear from you soon.\r\n\r\nThanks", "votes": null }, { "id": "97388", "postDate": "10/26/2015 20:24:15", "content": "Hi! It seems I have trouble trying to verify my account, so I'll contact you as soon as the support team resolves my issue. :-)
", "rawMarkdown": "Hi! It seems I have trouble trying to verify my account, so I'll contact you as soon as the support team resolves my issue. :-)", "votes": null }, { "id": "97389", "postDate": "10/26/2015 20:48:46", "content": "Sounds good
", "rawMarkdown": "Sounds good", "votes": null }, { "id": "97446", "postDate": "10/27/2015 16:51:50", "content": "Hey, i am interested in cv and machine learning. And I also use the theano and keras. Could I join your team and how could i contact you
", "rawMarkdown": "Hey, i am interested in cv and machine learning. And I also use the theano and keras. Could I join your team and how could i contact you", "votes": null }, { "id": "97490", "postDate": "10/27/2015 21:52:44", "content": "OK, now I feel really stupid, because somehow I thought that verifying my account will give me the access to the private messaging system, which apparently it does not. :-) So you can reach me by email (art1783@gmail.com) or Skype (tragedist) or Facebook (just look my name up).
", "rawMarkdown": "OK, now I feel really stupid, because somehow I thought that verifying my account will give me the access to the private messaging system, which apparently it does not. :-) So you can reach me by email (art1783@gmail.com) or Skype (tragedist) or Facebook (just look my name up).", "votes": null }, { "id": "98781", "postDate": "11/13/2015 14:57:33", "content": "@vbk, speaking about matlab and convnet...http://www.vlfeat.org/matconvnet/
", "rawMarkdown": "vbk, speaking about matlab and convnet...http://www.vlfeat.org/matconvnet/", "votes": null }, { "id": "98782", "postDate": "11/13/2015 15:43:52", "content": "Thanks @old-ufo
\n\nI already know about matconvnet. I use my own code (part of my own learning).
", "rawMarkdown": "Thanks @old-ufo\r\n\r\nI already know about matconvnet. I use my own code (part of my own learning).", "votes": null }, { "id": "99560", "postDate": "11/26/2015 13:39:06", "content": "I've summed up my progress so far in a blog post, just in case that maybe somebody'll find it useful.
", "rawMarkdown": "I've summed up my progress so far in a [blog post][1], just in case that maybe somebody'll find it useful.\r\n\r\n\r\n [1]: http://rocknrollnerd.github.io/ml/2015/11/26/right-whales.html", "votes": null } ], "comments": [ { "id": 93546, "author_name": "senecaur", "author_url": "", "post_date": "09/27/2015 19:14:05", "content": "[quote=Artem Khurshudov;93511]
\n\nPersonally I immediately rushed for a convolutional network, and surprisingly, ended up with nice enough looking results... although, of course, not perfect.
\n\n[/quote]
\n\nIf you don't mind saying, are you localizing using a convnet by regressing a bounding box or are you using a convnet as a classifier and doing some kind of sliding window approach? The results are pretty good! I am going to try the regression approach.
", "votes": null, "replies": [] }, { "id": 93548, "author_name": "rocknrollnerd", "author_url": "", "post_date": "09/27/2015 20:51:54", "content": "I use the sliding window approach, and actually I don't like it much - it still produces a lot of false detections, so I had to set a threshold for network's output value. And by "regressing a bounding box" do you mean something like Yann LeCun's OverFeat framework, or do you plan to make a network that directly outputs bounding box parameters? I thought about the latter myself, but not quite sure that'll work...
", "votes": null, "replies": [] }, { "id": 93550, "author_name": "senecaur", "author_url": "", "post_date": "09/27/2015 20:55:28", "content": "I mean more like the R-CNN approach. So getting candidate detections from some other method, e.g. cascades, and then simultaneously mitigating false alarms and regressing the bounding box using a CNN
", "votes": null, "replies": [] }, { "id": 93551, "author_name": "rocknrollnerd", "author_url": "", "post_date": "09/27/2015 20:59:23", "content": "Huh. Never heard of it before, so thanks, I guess I have a paper to read tonight. :-)
", "votes": null, "replies": [] }, { "id": 97367, "author_name": "rocknrollnerd", "author_url": "", "post_date": "10/26/2015 14:22:17", "content": "So, a little update on my approach, if anyone's interested: instead of using sliding windows/OverFeat/RCNN I focused on a mask regression technique described by Szegedy et al. Basically you make a binary black and white mask where everything inside the object's bounding box is white and everything else is black. Then you train the network to predict a mask from raw pixel image (obviously, you need to prepare a training set of masks). I spent almost 3 weeks trying to make make it work, and the key to success seems to be augmenting your training set 10x more and adding lots of dropout, and then even a fairly shallow convnet does the trick.
\n\nThen another idea I dwelled on for a while was to detect not the head specifically but the whole whale body instead, because I thought it will be much easier since heads are sometimes hardly distinguishable from the water-and-foam pattern. And here are some results: "mask1-2-3" displays mask prediction accuracy and how a bounded box is fitted to the predicted mask again (nothing fancy, some linear algebra and PCA), and "predicted_bbox-1-2-3" demonstrates the final result so far, separating whales from background.
\n\n...as a side note, anybody who isn't in the team yet and likes the mask regression approach, is interested to team up? I'm a PhD student in machine learning and computer vision, who uses python/Theano/Lasagne/Keras, and this is my first Kaggle competition, so not much experience so far.
", "votes": null, "replies": [] }, { "id": 97386, "author_name": "kolachalama", "author_url": "", "post_date": "10/26/2015 20:02:32", "content": "Artem,
\n\nThe results you got seem interesting. I'm interested to team up if you are open.
\n\nBasically, I'm trying to build a ConvNet using Matlab (I know this is a hard task, but not impossible). I have decent computing power to do some test runs on my end. On the side, I'm also trying to build a similar technique to isolate the whales but your method seems to produce better results.
\n\nWe can take this conversation private if you have any further questions. I hope to hear from you soon.
\n\nThanks
", "votes": null, "replies": [] }, { "id": 97388, "author_name": "rocknrollnerd", "author_url": "", "post_date": "10/26/2015 20:24:15", "content": "Hi! It seems I have trouble trying to verify my account, so I'll contact you as soon as the support team resolves my issue. :-)
", "votes": null, "replies": [] }, { "id": 97389, "author_name": "kolachalama", "author_url": "", "post_date": "10/26/2015 20:48:46", "content": "Sounds good
", "votes": null, "replies": [] }, { "id": 97446, "author_name": "wcskkk375", "author_url": "", "post_date": "10/27/2015 16:51:50", "content": "Hey, i am interested in cv and machine learning. And I also use the theano and keras. Could I join your team and how could i contact you
", "votes": null, "replies": [] }, { "id": 97490, "author_name": "rocknrollnerd", "author_url": "", "post_date": "10/27/2015 21:52:44", "content": "OK, now I feel really stupid, because somehow I thought that verifying my account will give me the access to the private messaging system, which apparently it does not. :-) So you can reach me by email (art1783@gmail.com) or Skype (tragedist) or Facebook (just look my name up).
", "votes": null, "replies": [] }, { "id": 98781, "author_name": "oldufo", "author_url": "", "post_date": "11/13/2015 14:57:33", "content": "@vbk, speaking about matlab and convnet...http://www.vlfeat.org/matconvnet/
", "votes": null, "replies": [] }, { "id": 98782, "author_name": "kolachalama", "author_url": "", "post_date": "11/13/2015 15:43:52", "content": "Thanks @old-ufo
\n\nI already know about matconvnet. I use my own code (part of my own learning).
", "votes": null, "replies": [] }, { "id": 99560, "author_name": "rocknrollnerd", "author_url": "", "post_date": "11/26/2015 13:39:06", "content": "I've summed up my progress so far in a blog post, just in case that maybe somebody'll find it useful.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "93511": "So I've been wondering if anybody's using techniques other than LBP model described by organizers. Personally I immediately rushed for a convolutional network, and surprisingly, ended up with nice enough looking results... although, of course, not perfect. The false positive cases have already been briefly discussed before, I guess - those situations where a head is obscured and some other body part (like a tail) is surrounded by foam and makes a distinct white-gray pattern. \r\n\r\nAnd then there's an idea - maybe extracting and processing local head/background patches just doesn't give you enough information. Maybe we have to consider the spatial stucture of the body first - and direct our attention at one of it's narrow ends. The body should be relatively easy to detect; and then it's only a matter of determining its orientation and distinguishing a head from a tail...", "93546": "[quote=Artem Khurshudov;93511]\r\n\r\nPersonally I immediately rushed for a convolutional network, and surprisingly, ended up with nice enough looking results... although, of course, not perfect. \r\n\r\n[/quote]\r\n\r\nIf you don't mind saying, are you localizing using a convnet by regressing a bounding box or are you using a convnet as a classifier and doing some kind of sliding window approach? The results are pretty good! I am going to try the regression approach.", "93548": "I use the sliding window approach, and actually I don't like it much - it still produces a lot of false detections, so I had to set a threshold for network's output value. And by \"regressing a bounding box\" do you mean something like Yann LeCun's [OverFeat][1] framework, or do you plan to make a network that directly outputs bounding box parameters? I thought about the latter myself, but not quite sure that'll work...\r\n\r\n\r\n [1]: http://arxiv.org/abs/1312.6229", "93550": "I mean more like the R-CNN approach. So getting candidate detections from some other method, e.g. cascades, and then simultaneously mitigating false alarms and regressing the bounding box using a CNN", "93551": "Huh. Never heard of it before, so thanks, I guess I have a paper to read tonight. :-)", "97367": "So, a little update on my approach, if anyone's interested: instead of using sliding windows/OverFeat/RCNN I focused on a mask regression technique described by [Szegedy et al][1]. Basically you make a binary black and white mask where everything inside the object's bounding box is white and everything else is black. Then you train the network to predict a mask from raw pixel image (obviously, you need to prepare a training set of masks). I spent almost 3 weeks trying to make make it work, and the key to success seems to be augmenting your training set 10x more and adding lots of dropout, and then even a fairly shallow convnet does the trick. \r\n\r\nThen another idea I dwelled on for a while was to detect not the head specifically but the whole whale body instead, because I thought it will be much easier since heads are sometimes hardly distinguishable from the water-and-foam pattern. And here are some results: \"mask1-2-3\" displays mask prediction accuracy and how a bounded box is fitted to the predicted mask again (nothing fancy, some linear algebra and PCA), and \"predicted_bbox-1-2-3\" demonstrates the final result so far, separating whales from background.\r\n\r\n...as a side note, anybody who isn't in the team yet and likes the mask regression approach, is interested to team up? I'm a PhD student in machine learning and computer vision, who uses python/Theano/Lasagne/Keras, and this is my first Kaggle competition, so not much experience so far.\r\n\r\n [1]: http://papers.nips.cc/paper/5207-deep-neural-networks-for-object-detection.pdf", "97386": "Artem, \r\n\r\nThe results you got seem interesting. I'm interested to team up if you are open.\r\n\r\nBasically, I'm trying to build a ConvNet using Matlab (I know this is a hard task, but not impossible). I have decent computing power to do some test runs on my end. On the side, I'm also trying to build a similar technique to isolate the whales but your method seems to produce better results.\r\n\r\nWe can take this conversation private if you have any further questions. I hope to hear from you soon.\r\n\r\nThanks", "97388": "Hi! It seems I have trouble trying to verify my account, so I'll contact you as soon as the support team resolves my issue. :-)", "97389": "Sounds good", "97446": "Hey, i am interested in cv and machine learning. And I also use the theano and keras. Could I join your team and how could i contact you", "97490": "OK, now I feel really stupid, because somehow I thought that verifying my account will give me the access to the private messaging system, which apparently it does not. :-) So you can reach me by email (art1783@gmail.com) or Skype (tragedist) or Facebook (just look my name up).", "98781": "vbk, speaking about matlab and convnet...http://www.vlfeat.org/matconvnet/", "98782": "Thanks @old-ufo\r\n\r\nI already know about matconvnet. I use my own code (part of my own learning).", "99560": "I've summed up my progress so far in a [blog post][1], just in case that maybe somebody'll find it useful.\r\n\r\n\r\n [1]: http://rocknrollnerd.github.io/ml/2015/11/26/right-whales.html" }, "source": "meta" }