Do I have to train a detector to detect the whale head? Maybe I can crop the whale head by hand.
", "messages": [ { "id": "98973", "postDate": "11/17/2015 13:37:16", "content": "Do I have to train a detector to detect the whale head? Maybe I can crop the whale head by hand.
", "rawMarkdown": "Do I have to train a detector to detect the whale head? Maybe I can crop the whale head by hand.", "votes": null }, { "id": "98975", "postDate": "11/17/2015 14:22:29", "content": "For the training images, yes, you could hand label them. But the problem comes when your whale identifier is deployed in the field - you aren't there to crop the image!
\n\nOne could build a system that is a combination of human and machine - the human crops (and maybe additionally points out the salient features), the machine identifies. And in some applications that makes a lot of sense. But the researchers have asked for a system that just looks at an image and provides an ID with no human intervention. So you'll need to come up with a way to automatically crop the images in the test set, just like it would work in the field.
", "rawMarkdown": "For the training images, yes, you could hand label them. But the problem comes when your whale identifier is deployed in the field - you aren't there to crop the image! \r\n\r\nOne could build a system that is a combination of human and machine - the human crops (and maybe additionally points out the salient features), the machine identifies. And in some applications that makes a lot of sense. But the researchers have asked for a system that just looks at an image and provides an ID with no human intervention. So you'll need to come up with a way to automatically crop the images in the test set, just like it would work in the field.", "votes": null }, { "id": "98994", "postDate": "11/17/2015 22:20:37", "content": "Good points there @john. I think a hybrid human-computer system will provide the greatest benefit.
\n\nThe first step of cropping the head is rather trivial for any human, even without any training. This can be done effortlessly with virtually 100% accuracy.\nWhale identification on the other hand is challenging for both human and computer.
\n\nA hybrid system with human aid gives the computer a better shot at the identification task. \nThis might be a more practical setting.
", "rawMarkdown": "Good points there @john. I think a hybrid human-computer system will provide the greatest benefit. \r\n\r\nThe first step of cropping the head is rather trivial for any human, even without any training. This can be done effortlessly with virtually 100% accuracy.\r\nWhale identification on the other hand is challenging for both human and computer.\r\n\r\nA hybrid system with human aid gives the computer a better shot at the identification task. \r\nThis might be a more practical setting.", "votes": null }, { "id": "98997", "postDate": "11/17/2015 22:47:06", "content": "[quote=Vinh Nguyen;98994]\nA hybrid system with human aid gives the computer a better shot at the identification task. \nThis might be a more practical setting.\n[/quote]
\n\nWell you couldn't win the competition with one, since the rules are specific, but if you made a model like that where it only took seconds to crop per image (and outperformed everyone else on the leaderboards), then the sponsors might be interested in your work.
", "rawMarkdown": "[quote=Vinh Nguyen;98994]\r\nA hybrid system with human aid gives the computer a better shot at the identification task. \r\nThis might be a more practical setting.\r\n[/quote]\r\n\r\nWell you couldn't win the competition with one, since the rules are specific, but if you made a model like that where it only took seconds to crop per image (and outperformed everyone else on the leaderboards), then the sponsors might be interested in your work.", "votes": null }, { "id": "98998", "postDate": "11/17/2015 23:28:53", "content": "Indeed, I think it is the sponsor's top interest to build a practically useful system, which could be like this:
\n\nA 100% automatic solution would certainly be nicer, but a little effortless intervention for any improved performance is not such a bad idea.
\n\nThis is a challenging problem and I don't think the current solutions are anywhere near satisfactory for field deployment. But let's see...
", "rawMarkdown": "Indeed, I think it is the sponsor's top interest to build a practically useful system, which could be like this:\r\n\r\n- The operator takes a photo with an Iphone or tablet app\r\n- The app suggests a crop area for whale face\r\n- The operator either approves, or manually corrects the crop\r\n- The app makes the final prediction\r\n\r\nA 100% automatic solution would certainly be nicer, but a little effortless intervention for any improved performance is not such a bad idea. \r\n\r\nThis is a challenging problem and I don't think the current solutions are anywhere near satisfactory for field deployment. But let's see...", "votes": null }, { "id": "99002", "postDate": "11/18/2015 00:38:35", "content": "@Vinh Nguyen, I agree with you. I think that the compition can be separate to two tasks.The first is the whale head detection like hunman face detection .The second is the whale id identification. As the first task is so difficult.If we can't detect the whale head, we will can't the identificate the whale id. If we see a whale at any where, wo can take a picture and crop the head . We do like that, the whale id can be get more acurrate.
", "rawMarkdown": "Vinh Nguyen, I agree with you. I think that the compition can be separate to two tasks.The first is the whale head detection like hunman face detection .The second is the whale id identification. As the first task is so difficult.If we can't detect the whale head, we will can't the identificate the whale id. If we see a whale at any where, wo can take a picture and crop the head . We do like that, the whale id can be get more acurrate.", "votes": null } ], "comments": [ { "id": 98975, "author_name": "johnblankenbaker", "author_url": "", "post_date": "11/17/2015 14:22:29", "content": "For the training images, yes, you could hand label them. But the problem comes when your whale identifier is deployed in the field - you aren't there to crop the image!
\n\nOne could build a system that is a combination of human and machine - the human crops (and maybe additionally points out the salient features), the machine identifies. And in some applications that makes a lot of sense. But the researchers have asked for a system that just looks at an image and provides an ID with no human intervention. So you'll need to come up with a way to automatically crop the images in the test set, just like it would work in the field.
", "votes": null, "replies": [] }, { "id": 98994, "author_name": "vinhnguyen", "author_url": "", "post_date": "11/17/2015 22:20:37", "content": "Good points there @john. I think a hybrid human-computer system will provide the greatest benefit.
\n\nThe first step of cropping the head is rather trivial for any human, even without any training. This can be done effortlessly with virtually 100% accuracy.\nWhale identification on the other hand is challenging for both human and computer.
\n\nA hybrid system with human aid gives the computer a better shot at the identification task. \nThis might be a more practical setting.
", "votes": null, "replies": [] }, { "id": 98997, "author_name": "piinformatics", "author_url": "", "post_date": "11/17/2015 22:47:06", "content": "[quote=Vinh Nguyen;98994]\nA hybrid system with human aid gives the computer a better shot at the identification task. \nThis might be a more practical setting.\n[/quote]
\n\nWell you couldn't win the competition with one, since the rules are specific, but if you made a model like that where it only took seconds to crop per image (and outperformed everyone else on the leaderboards), then the sponsors might be interested in your work.
", "votes": null, "replies": [] }, { "id": 98998, "author_name": "vinhnguyen", "author_url": "", "post_date": "11/17/2015 23:28:53", "content": "Indeed, I think it is the sponsor's top interest to build a practically useful system, which could be like this:
\n\nA 100% automatic solution would certainly be nicer, but a little effortless intervention for any improved performance is not such a bad idea.
\n\nThis is a challenging problem and I don't think the current solutions are anywhere near satisfactory for field deployment. But let's see...
", "votes": null, "replies": [] }, { "id": 99002, "author_name": "wuqiangchiq", "author_url": "", "post_date": "11/18/2015 00:38:35", "content": "@Vinh Nguyen, I agree with you. I think that the compition can be separate to two tasks.The first is the whale head detection like hunman face detection .The second is the whale id identification. As the first task is so difficult.If we can't detect the whale head, we will can't the identificate the whale id. If we see a whale at any where, wo can take a picture and crop the head . We do like that, the whale id can be get more acurrate.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "98973": "Do I have to train a detector to detect the whale head? Maybe I can crop the whale head by hand.", "98975": "For the training images, yes, you could hand label them. But the problem comes when your whale identifier is deployed in the field - you aren't there to crop the image! \r\n\r\nOne could build a system that is a combination of human and machine - the human crops (and maybe additionally points out the salient features), the machine identifies. And in some applications that makes a lot of sense. But the researchers have asked for a system that just looks at an image and provides an ID with no human intervention. So you'll need to come up with a way to automatically crop the images in the test set, just like it would work in the field.", "98994": "Good points there @john. I think a hybrid human-computer system will provide the greatest benefit. \r\n\r\nThe first step of cropping the head is rather trivial for any human, even without any training. This can be done effortlessly with virtually 100% accuracy.\r\nWhale identification on the other hand is challenging for both human and computer.\r\n\r\nA hybrid system with human aid gives the computer a better shot at the identification task. \r\nThis might be a more practical setting.", "98997": "[quote=Vinh Nguyen;98994]\r\nA hybrid system with human aid gives the computer a better shot at the identification task. \r\nThis might be a more practical setting.\r\n[/quote]\r\n\r\nWell you couldn't win the competition with one, since the rules are specific, but if you made a model like that where it only took seconds to crop per image (and outperformed everyone else on the leaderboards), then the sponsors might be interested in your work.", "98998": "Indeed, I think it is the sponsor's top interest to build a practically useful system, which could be like this:\r\n\r\n- The operator takes a photo with an Iphone or tablet app\r\n- The app suggests a crop area for whale face\r\n- The operator either approves, or manually corrects the crop\r\n- The app makes the final prediction\r\n\r\nA 100% automatic solution would certainly be nicer, but a little effortless intervention for any improved performance is not such a bad idea. \r\n\r\nThis is a challenging problem and I don't think the current solutions are anywhere near satisfactory for field deployment. But let's see...", "99002": "Vinh Nguyen, I agree with you. I think that the compition can be separate to two tasks.The first is the whale head detection like hunman face detection .The second is the whale id identification. As the first task is so difficult.If we can't detect the whale head, we will can't the identificate the whale id. If we see a whale at any where, wo can take a picture and crop the head . We do like that, the whale id can be get more acurrate." }, "source": "meta" }