Hello,
\n\nI have been trying to implement some unsupervised whale detection. The main idea is that in the pictures there are potentially only the whale and water so using the ratio of color R/B and R/G to detect the whale.
\n\nIt work well it many cases but I don't know what is the real accuracy since I didn't do the manual labelling. It would be great if any one who did could share with me the evaluation ;)
\n\nAny idea and comment are welcome!
", "messages": [ { "id": "93498", "postDate": "09/26/2015 17:50:48", "content": "Hello,
\n\nI have been trying to implement some unsupervised whale detection. The main idea is that in the pictures there are potentially only the whale and water so using the ratio of color R/B and R/G to detect the whale.
\n\nIt work well it many cases but I don't know what is the real accuracy since I didn't do the manual labelling. It would be great if any one who did could share with me the evaluation ;)
\n\nAny idea and comment are welcome!
", "rawMarkdown": "Hello,\r\n\r\nI have been trying to implement some unsupervised whale detection. The main idea is that in the pictures there are potentially only the whale and water so using the ratio of color R/B and R/G to detect the whale.\r\n\r\nIt work well it many cases but I don't know what is the real accuracy since I didn't do the manual labelling. It would be great if any one who did could share with me the evaluation ;) \r\n\r\nAny idea and comment are welcome!", "votes": null }, { "id": "93529", "postDate": "09/27/2015 12:16:46", "content": "Hey,
\n\nI tried your code on all the images, it gave up on w_8512.jpg.
\n\nAlso, checked randomly, w_965, w_706 , w_1084 was not correctly detected.
\n\nThanks\nManish
", "rawMarkdown": "Hey,\r\n\r\nI tried your code on all the images, it gave up on w_8512.jpg.\r\n\r\nAlso, checked randomly, w_965, w_706 , w_1084 was not correctly detected.\r\n\r\nThanks\r\nManish", "votes": null }, { "id": "93628", "postDate": "09/29/2015 10:58:27", "content": "There are better image processing techniques, like edge detector , check that out.
", "rawMarkdown": "There are better image processing techniques, like edge detector , check that out.", "votes": null }, { "id": "95221", "postDate": "10/06/2015 03:25:55", "content": "[quote=SecondPlan;93628]
\n\nThere are better image processing techniques, like edge detector , check that out.
\n\n[/quote]
\n\nOpenCV has a Canny Edge detection algorithm that might be a good place to start if you decide to go this route. Easy to use, but OpenCV can be a bit of a bear to install and configure.
", "rawMarkdown": "[quote=SecondPlan;93628]\r\n\r\nThere are better image processing techniques, like edge detector , check that out.\r\n\r\n[/quote]\r\n\r\nOpenCV has a Canny Edge detection algorithm that might be a good place to start if you decide to go this route. Easy to use, but OpenCV can be a bit of a bear to install and configure.", "votes": null }, { "id": "95562", "postDate": "10/09/2015 08:09:15", "content": "The idea of separating the sea and the whale is effective. But its efficiency of about 70 percent.
\n\nI have not turned the division on the basis of segmentation (meanshit) and separation of the ribs (Canny Edge). More precisely I tried, but has not achieved any clear and accurate results. I walked some other way, the separation of the similarity of the squares.
\n\nHere are some of the results. Maybe it will help someone. The csv file results evaluation of a square where the whale (first 1000 files). The square is contained in the fields "r0,c0,r1,c1", the field "pie" contains the number of pieces that managed to leave. Where 0 or greater than 1, the result is poor. Examples also joined.
\n\nThe data so far are intermediate.
", "rawMarkdown": "The idea of separating the sea and the whale is effective. But its efficiency of about 70 percent. \r\n\r\nI have not turned the division on the basis of segmentation (meanshit) and separation of the ribs (Canny Edge). More precisely I tried, but has not achieved any clear and accurate results. I walked some other way, the separation of the similarity of the squares. \r\n\r\nHere are some of the results. Maybe it will help someone. The csv file results evaluation of a square where the whale (first 1000 files). The square is contained in the fields \"r0,c0,r1,c1\", the field \"pie\" contains the number of pieces that managed to leave. Where 0 or greater than 1, the result is poor. Examples also joined.\r\n\r\nThe data so far are intermediate.", "votes": null }, { "id": "95759", "postDate": "10/11/2015 15:06:30", "content": "Hi all,\nThanks for your feedbacks. I realized that my approach was quite "naive". Actually the most advanced method for object detection is Selective Search (http://koen.me/research/selectivesearch/). This should be much more complete approach than I can ever think of :)
\n\nHowever, if I understand the paper correctly, this method would yield regions containing the whole whale. If you want to just detect the head (like me), I think an supervised approach would be better.
\n\nFinally, I went for a sliding windows + CNN approach using the sloth annotations put together by Smerity and some other participants (https://www.kaggle.com/c/noaa-right-whale-recognition/forums/t/16353/annotated-faces-for-noaa-right-whale) Big thanks to them
", "rawMarkdown": "Hi all,\r\nThanks for your feedbacks. I realized that my approach was quite \"naive\". Actually the most advanced method for object detection is Selective Search (http://koen.me/research/selectivesearch/). This should be much more complete approach than I can ever think of :)\r\n\r\nHowever, if I understand the paper correctly, this method would yield regions containing the whole whale. If you want to just detect the head (like me), I think an supervised approach would be better.\r\n\r\nFinally, I went for a sliding windows + CNN approach using the sloth annotations put together by Smerity and some other participants (https://www.kaggle.com/c/noaa-right-whale-recognition/forums/t/16353/annotated-faces-for-noaa-right-whale) Big thanks to them", "votes": null } ], "comments": [ { "id": 93529, "author_name": "manishap", "author_url": "", "post_date": "09/27/2015 12:16:46", "content": "Hey,
\n\nI tried your code on all the images, it gave up on w_8512.jpg.
\n\nAlso, checked randomly, w_965, w_706 , w_1084 was not correctly detected.
\n\nThanks\nManish
", "votes": null, "replies": [] }, { "id": 93628, "author_name": "usixuz", "author_url": "", "post_date": "09/29/2015 10:58:27", "content": "There are better image processing techniques, like edge detector , check that out.
", "votes": null, "replies": [] }, { "id": 95221, "author_name": "", "author_url": "", "post_date": "10/06/2015 03:25:55", "content": "[quote=SecondPlan;93628]
\n\nThere are better image processing techniques, like edge detector , check that out.
\n\n[/quote]
\n\nOpenCV has a Canny Edge detection algorithm that might be a good place to start if you decide to go this route. Easy to use, but OpenCV can be a bit of a bear to install and configure.
", "votes": null, "replies": [] }, { "id": 95562, "author_name": "sorokinv", "author_url": "", "post_date": "10/09/2015 08:09:15", "content": "The idea of separating the sea and the whale is effective. But its efficiency of about 70 percent.
\n\nI have not turned the division on the basis of segmentation (meanshit) and separation of the ribs (Canny Edge). More precisely I tried, but has not achieved any clear and accurate results. I walked some other way, the separation of the similarity of the squares.
\n\nHere are some of the results. Maybe it will help someone. The csv file results evaluation of a square where the whale (first 1000 files). The square is contained in the fields "r0,c0,r1,c1", the field "pie" contains the number of pieces that managed to leave. Where 0 or greater than 1, the result is poor. Examples also joined.
\n\nThe data so far are intermediate.
", "votes": null, "replies": [] }, { "id": 95759, "author_name": "lamdang", "author_url": "", "post_date": "10/11/2015 15:06:30", "content": "Hi all,\nThanks for your feedbacks. I realized that my approach was quite "naive". Actually the most advanced method for object detection is Selective Search (http://koen.me/research/selectivesearch/). This should be much more complete approach than I can ever think of :)
\n\nHowever, if I understand the paper correctly, this method would yield regions containing the whole whale. If you want to just detect the head (like me), I think an supervised approach would be better.
\n\nFinally, I went for a sliding windows + CNN approach using the sloth annotations put together by Smerity and some other participants (https://www.kaggle.com/c/noaa-right-whale-recognition/forums/t/16353/annotated-faces-for-noaa-right-whale) Big thanks to them
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "93498": "Hello,\r\n\r\nI have been trying to implement some unsupervised whale detection. The main idea is that in the pictures there are potentially only the whale and water so using the ratio of color R/B and R/G to detect the whale.\r\n\r\nIt work well it many cases but I don't know what is the real accuracy since I didn't do the manual labelling. It would be great if any one who did could share with me the evaluation ;) \r\n\r\nAny idea and comment are welcome!", "93529": "Hey,\r\n\r\nI tried your code on all the images, it gave up on w_8512.jpg.\r\n\r\nAlso, checked randomly, w_965, w_706 , w_1084 was not correctly detected.\r\n\r\nThanks\r\nManish", "93628": "There are better image processing techniques, like edge detector , check that out.", "95221": "[quote=SecondPlan;93628]\r\n\r\nThere are better image processing techniques, like edge detector , check that out.\r\n\r\n[/quote]\r\n\r\nOpenCV has a Canny Edge detection algorithm that might be a good place to start if you decide to go this route. Easy to use, but OpenCV can be a bit of a bear to install and configure.", "95562": "The idea of separating the sea and the whale is effective. But its efficiency of about 70 percent. \r\n\r\nI have not turned the division on the basis of segmentation (meanshit) and separation of the ribs (Canny Edge). More precisely I tried, but has not achieved any clear and accurate results. I walked some other way, the separation of the similarity of the squares. \r\n\r\nHere are some of the results. Maybe it will help someone. The csv file results evaluation of a square where the whale (first 1000 files). The square is contained in the fields \"r0,c0,r1,c1\", the field \"pie\" contains the number of pieces that managed to leave. Where 0 or greater than 1, the result is poor. Examples also joined.\r\n\r\nThe data so far are intermediate.", "95759": "Hi all,\r\nThanks for your feedbacks. I realized that my approach was quite \"naive\". Actually the most advanced method for object detection is Selective Search (http://koen.me/research/selectivesearch/). This should be much more complete approach than I can ever think of :)\r\n\r\nHowever, if I understand the paper correctly, this method would yield regions containing the whole whale. If you want to just detect the head (like me), I think an supervised approach would be better.\r\n\r\nFinally, I went for a sliding windows + CNN approach using the sloth annotations put together by Smerity and some other participants (https://www.kaggle.com/c/noaa-right-whale-recognition/forums/t/16353/annotated-faces-for-noaa-right-whale) Big thanks to them" }, "source": "meta" }