Seems to me, first we need to be able to identify the 17 body zones well in any image and then learn using the training data which zones are anomalous..
\n\nSo we have training data learn from for anomaly detection but no training data for body zone learning. We are on our own here for that right ?
", "messages": [ { "id": 195536, "postDate": "2017-06-23T22:57:50.240Z", "content": "Seems to me, first we need to be able to identify the 17 body zones well in any image and then learn using the training data which zones are anomalous..
\n\nSo we have training data learn from for anomaly detection but no training data for body zone learning. We are on our own here for that right ?
", "rawMarkdown": "Seems to me, first we need to be able to identify the 17 body zones well in any image and then learn using the training data which zones are anomalous.. \n\nSo we have training data learn from for anomaly detection but no training data for body zone learning. We are on our own here for that right ? ", "votes": 8 }, { "id": 195574, "postDate": "2017-06-24T03:11:25.690Z", "content": "That is how I have read it as well. Basically, we would need to figure out where on the image(s) the zones are located. Then, for each zone, anomaly detection would need to be performed.
\n\nHowever, another option could be the opposite approach where you detect if there is an anomaly and then figure out which zone(s) it is in. I have been toying with both approaches.
", "rawMarkdown": "That is how I have read it as well. Basically, we would need to figure out where on the image(s) the zones are located. Then, for each zone, anomaly detection would need to be performed. \n\nHowever, another option could be the opposite approach where you detect if there is an anomaly and then figure out which zone(s) it is in. I have been toying with both approaches.", "votes": 3 }, { "id": 197717, "postDate": "2017-06-30T04:05:05.487Z", "content": "A third possibility. Treat it not as a binary classification problem but a multinomial prediction problem, meaning for each image there are 17 potential outcomes whose probability sums to 1. Then you don't have go around identifying zones in a image - just give the learning algorithm the whole set of data and ask it to learn and predict the 17 outcomes. I've done that using tensorflow and a simple back-propagation NN - it wasn't a very good result the first time ! - but it shakes out the overall code quite well.
", "rawMarkdown": "A third possibility. Treat it not as a binary classification problem but a multinomial prediction problem, meaning for each image there are 17 potential outcomes whose probability sums to 1. Then you don't have go around identifying zones in a image - just give the learning algorithm the whole set of data and ask it to learn and predict the 17 outcomes. I've done that using tensorflow and a simple back-propagation NN - it wasn't a very good result the first time ! - but it shakes out the overall code quite well. ", "votes": 2, "replies": [ { "id": 199118, "postDate": "2017-07-04T16:49:06.087Z", "content": "Some images have more than one threat.
\n\nI think the dataset is too small for a model to learn to detect the threats and the locations at the same time.
", "rawMarkdown": "Some images have more than one threat.\n\nI think the dataset is too small for a model to learn to detect the threats and the locations at the same time. ", "votes": 1 } ] }, { "id": 205313, "postDate": "2017-07-21T03:18:42.140Z", "content": "I am newbie in machine learning as well as Kaggle; can we do image segmentation ( in required 16 or 17 segments) by clustering like K means? And then go train CNN with each segment ?
", "rawMarkdown": "I am newbie in machine learning as well as Kaggle; can we do image segmentation ( in required 16 or 17 segments) by clustering like K means? And then go train CNN with each segment ?" }, { "id": 195608, "postDate": "2017-06-24T07:30:52.203Z", "rawMarkdown": "", "votes": 3, "isDeleted": true } ], "comments": [ { "id": 195574, "author_name": "Patrick McNeil", "author_url": "", "post_date": "2017-06-24T03:11:25.690000", "content": "That is how I have read it as well. Basically, we would need to figure out where on the image(s) the zones are located. Then, for each zone, anomaly detection would need to be performed.
\n\nHowever, another option could be the opposite approach where you detect if there is an anomaly and then figure out which zone(s) it is in. I have been toying with both approaches.
", "votes": 3, "replies": [] }, { "id": 197717, "author_name": "skthetwo", "author_url": "", "post_date": "2017-06-30T04:05:05.487000", "content": "A third possibility. Treat it not as a binary classification problem but a multinomial prediction problem, meaning for each image there are 17 potential outcomes whose probability sums to 1. Then you don't have go around identifying zones in a image - just give the learning algorithm the whole set of data and ask it to learn and predict the 17 outcomes. I've done that using tensorflow and a simple back-propagation NN - it wasn't a very good result the first time ! - but it shakes out the overall code quite well.
", "votes": 2, "replies": [ { "id": 199118, "author_name": "Guillermo Barbadillo", "author_url": "", "post_date": "2017-07-04T16:49:06.087000", "content": "Some images have more than one threat.
\n\nI think the dataset is too small for a model to learn to detect the threats and the locations at the same time.
", "votes": 1, "replies": [] } ] }, { "id": 205313, "author_name": "ThinkBig", "author_url": "", "post_date": "2017-07-21T03:18:42.140000", "content": "I am newbie in machine learning as well as Kaggle; can we do image segmentation ( in required 16 or 17 segments) by clustering like K means? And then go train CNN with each segment ?
", "votes": 0, "replies": [] }, { "id": 195608, "author_name": "", "author_url": "", "post_date": "2017-06-24T07:30:52.203000", "content": "", "votes": 3, "replies": [] } ], "raw_markdown_by_id": { "195536": "Seems to me, first we need to be able to identify the 17 body zones well in any image and then learn using the training data which zones are anomalous.. \n\nSo we have training data learn from for anomaly detection but no training data for body zone learning. We are on our own here for that right ? ", "195574": "That is how I have read it as well. Basically, we would need to figure out where on the image(s) the zones are located. Then, for each zone, anomaly detection would need to be performed. \n\nHowever, another option could be the opposite approach where you detect if there is an anomaly and then figure out which zone(s) it is in. I have been toying with both approaches.", "197717": "A third possibility. Treat it not as a binary classification problem but a multinomial prediction problem, meaning for each image there are 17 potential outcomes whose probability sums to 1. Then you don't have go around identifying zones in a image - just give the learning algorithm the whole set of data and ask it to learn and predict the 17 outcomes. I've done that using tensorflow and a simple back-propagation NN - it wasn't a very good result the first time ! - but it shakes out the overall code quite well. ", "205313": "I am newbie in machine learning as well as Kaggle; can we do image segmentation ( in required 16 or 17 segments) by clustering like K means? And then go train CNN with each segment ?", "195608": "" } }