It seems like most of the examples have only one class_id. Is this class_id a category label?
\n\nIf the example has more than one class_id (ie. '10_3_20_33_60_70_91'), do we assume the first one is a category label and the rest are attributes?
\n\nIt's a little confusing because the category and attribute labels overlap (they both start at index 0).
", "messages": [ { "id": "523116", "postDate": "04/25/2019 15:39:47", "content": "It seems like most of the examples have only one class_id. Is this class_id a category label?
\n\nIf the example has more than one class_id (ie. '10_3_20_33_60_70_91'), do we assume the first one is a category label and the rest are attributes?
\n\nIt's a little confusing because the category and attribute labels overlap (they both start at index 0).
", "rawMarkdown": "It seems like most of the examples have only one class_id. Is this class_id a category label? \n\nIf the example has more than one class_id (ie. '10_3_20_33_60_70_91'), do we assume the first one is a category label and the rest are attributes? \n\nIt's a little confusing because the category and attribute labels overlap (they both start at index 0).", "votes": null }, { "id": "523135", "postDate": "04/25/2019 16:14:57", "content": "You are right, most of the training images (~80%) have one ClassId, which represents the main apparel category, like \"jacket\", or \"sleeve\", or \"glasses\". Out of those images that have attributes, only categories that are \"main apparel categories\"(category id from 0 - 12), like jacket, dress, pants, have attributes.
The first integer is a category label. And from second integer onwards, all labels are attributes. Both category and attribute are individually indexed.
", "rawMarkdown": "You are right, most of the training images (~80%) have one `ClassId`, which represents the main apparel category, like \"jacket\", or \"sleeve\", or \"glasses\". Out of those images that have attributes, only categories that are \"main apparel categories\"(category id from 0 - 12), like jacket, dress, pants, have attributes.\n\nThe first integer *is* a category label. And from second integer onwards, all labels are attributes. Both category and attribute are individually indexed.", "votes": null }, { "id": "523146", "postDate": "04/25/2019 16:40:35", "content": "Thanks!
", "rawMarkdown": "Thanks!", "votes": null }, { "id": "525084", "postDate": "04/30/2019 06:44:49", "content": "@makeitworkjml a related question, we should interpret those images that don't have attributes as images where attributes are not labelled, but could potentially be present, right? So in our ideal test submission we should have predicted attributes for such images (if they were in test)?
", "rawMarkdown": "makeitworkjml a related question, we should interpret those images that don't have attributes as images where attributes are not labelled, but could potentially be present, right? So in our ideal test submission we should have predicted attributes for such images (if they were in test)?", "votes": null }, { "id": "525231", "postDate": "04/30/2019 13:40:03", "content": "@lopuhin you are correct. In the test submission, all the masks in the test set were fully labeled.
", "rawMarkdown": "lopuhin you are correct. In the test submission, all the masks in the test set were fully labeled.", "votes": null }, { "id": "526271", "postDate": "05/02/2019 17:48:11", "content": "According to the description, the attributes are concatenated with the class ID. How exactly is this done? I was thinking if a particular image has a category and its attributes, its classID would be labelled as 'category, attribute1, attribute2' but that doesn't seem to be the case.
", "rawMarkdown": "According to the description, the attributes are concatenated with the class ID. How exactly is this done? I was thinking if a particular image has a category and its attributes, its classID would be labelled as 'category, attribute1, attribute2' but that doesn't seem to be the case.", "votes": null }, { "id": "526785", "postDate": "05/03/2019 18:41:44", "content": "Hi Sidhant, in this competition, we concatenated with \"_\", instead of \", \". But the basic idea is the same. \"category1_attribute12_attribute30\" means this masks has category 1, and attributes 12 and 30.
", "rawMarkdown": "Hi Sidhant, in this competition, we concatenated with \"_\", instead of \", \". But the basic idea is the same. \"category1_attribute12_attribute30\" means this masks has category 1, and attributes 12 and 30.", "votes": null }, { "id": "527940", "postDate": "05/06/2019 16:53:08", "content": "@makeitworkjml I have a related question - how to determine if attributes are labelled or not in the train set? Can we assume that if there is at least one attribute present in an image, then all objects in this image have attributes annotated, even if they are empty for some objects? Can there be objects in the training set that were annotated for attributes, and resulting attribute annotation turned out to be empty? If yes, can you provide us with a list of image ids where attributes were annotated, or some other way to find them?
", "rawMarkdown": "makeitworkjml I have a related question - how to determine if attributes are labelled or not in the train set? Can we assume that if there is at least one attribute present in an image, then all objects in this image have attributes annotated, even if they are empty for some objects? Can there be objects in the training set that were annotated for attributes, and resulting attribute annotation turned out to be empty? If yes, can you provide us with a list of image ids where attributes were annotated, or some other way to find them?", "votes": null }, { "id": "528053", "postDate": "05/07/2019 01:27:56", "content": "hi @lopuhin, we made sure that for images that do have attributes, only category id from 0 - 12 (main apparel categories) have attributes. \nImage with ClassId from \"0\" to \"12\" means that this image did not get annotated with fine-grained attributes.
Hope this helps!
", "rawMarkdown": "hi @lopuhin, we made sure that for images that do have attributes, only category id from 0 - 12 (main apparel categories) have attributes. \nImage with `ClassId` from \"0\" to \"12\" means that this image did not get annotated with fine-grained attributes.\n\nHope this helps!", "votes": null }, { "id": "528348", "postDate": "05/07/2019 14:44:40", "content": "Aha, thanks for clarification @makeitworkjml this makes sense.\nFor anyone who might think that two sentences are contradictory, like me at first, pay attention that category is not the same as ClassId, and ClassId \"0\" means that not attributes are present :)
", "rawMarkdown": "Aha, thanks for clarification @makeitworkjml this makes sense.\nFor anyone who might think that two sentences are contradictory, like me at first, pay attention that category is not the same as ClassId, and ClassId \"0\" means that not attributes are present :)", "votes": null }, { "id": "528706", "postDate": "05/08/2019 12:22:15", "content": "Hi Menglin,\ncould you comment regarding the evaluation with the attributes? Do we have to get a category and all attributes right in order to get a count for the segmented object? (to be clear: if I miss one attribute, but all the rest is correct and IoU is 0.8 it will not count at all)
", "rawMarkdown": "Hi Menglin,\ncould you comment regarding the evaluation with the attributes? Do we have to get a category and all attributes right in order to get a count for the segmented object? (to be clear: if I miss one attribute, but all the rest is correct and IoU is 0.8 it will not count at all)", "votes": null }, { "id": "529499", "postDate": "05/10/2019 03:42:40", "content": "hi @makeitworkjml , sorry I'm still a little bit confused. According to my statistics, there are also some categories annotated with attributes labels. If I don't make mistakes, class 27, 28 and 33 all contain one image annotated with attributes. Plus, I'm also curious whether there exists some category specific attribute. For example, there are only 14 instances belonging to cat12, where only 23 kinds of attributes appear in the cat12 training data. Should we assume that only these 23 attributes are used to describe cat12? Or other attributes can also be predicted for cat12 instances?
", "rawMarkdown": "hi @makeitworkjml , sorry I'm still a little bit confused. According to my statistics, there are also some categories annotated with attributes labels. If I don't make mistakes, class 27, 28 and 33 all contain one image annotated with attributes. Plus, I'm also curious whether there exists some category specific attribute. For example, there are only 14 instances belonging to cat12, where only 23 kinds of attributes appear in the cat12 training data. Should we assume that only these 23 attributes are used to describe cat12? Or other attributes can also be predicted for cat12 instances?", "votes": null } ], "comments": [ { "id": 523135, "author_name": "makeitworkjml", "author_url": "", "post_date": "04/25/2019 16:14:57", "content": "You are right, most of the training images (~80%) have one ClassId, which represents the main apparel category, like \"jacket\", or \"sleeve\", or \"glasses\". Out of those images that have attributes, only categories that are \"main apparel categories\"(category id from 0 - 12), like jacket, dress, pants, have attributes.
The first integer is a category label. And from second integer onwards, all labels are attributes. Both category and attribute are individually indexed.
", "votes": null, "replies": [ { "id": 523146, "author_name": "shaayaansayed", "author_url": "", "post_date": "04/25/2019 16:40:35", "content": "Thanks!
", "votes": null, "replies": [] }, { "id": 525084, "author_name": "lopuhin", "author_url": "", "post_date": "04/30/2019 06:44:49", "content": "@makeitworkjml a related question, we should interpret those images that don't have attributes as images where attributes are not labelled, but could potentially be present, right? So in our ideal test submission we should have predicted attributes for such images (if they were in test)?
", "votes": null, "replies": [] }, { "id": 525231, "author_name": "makeitworkjml", "author_url": "", "post_date": "04/30/2019 13:40:03", "content": "@lopuhin you are correct. In the test submission, all the masks in the test set were fully labeled.
", "votes": null, "replies": [] }, { "id": 526271, "author_name": "thedocs", "author_url": "", "post_date": "05/02/2019 17:48:11", "content": "According to the description, the attributes are concatenated with the class ID. How exactly is this done? I was thinking if a particular image has a category and its attributes, its classID would be labelled as 'category, attribute1, attribute2' but that doesn't seem to be the case.
", "votes": null, "replies": [] }, { "id": 526785, "author_name": "makeitworkjml", "author_url": "", "post_date": "05/03/2019 18:41:44", "content": "Hi Sidhant, in this competition, we concatenated with \"_\", instead of \", \". But the basic idea is the same. \"category1_attribute12_attribute30\" means this masks has category 1, and attributes 12 and 30.
", "votes": null, "replies": [] }, { "id": 528706, "author_name": "blondinka", "author_url": "", "post_date": "05/08/2019 12:22:15", "content": "Hi Menglin,\ncould you comment regarding the evaluation with the attributes? Do we have to get a category and all attributes right in order to get a count for the segmented object? (to be clear: if I miss one attribute, but all the rest is correct and IoU is 0.8 it will not count at all)
", "votes": null, "replies": [] } ] }, { "id": 527940, "author_name": "lopuhin", "author_url": "", "post_date": "05/06/2019 16:53:08", "content": "@makeitworkjml I have a related question - how to determine if attributes are labelled or not in the train set? Can we assume that if there is at least one attribute present in an image, then all objects in this image have attributes annotated, even if they are empty for some objects? Can there be objects in the training set that were annotated for attributes, and resulting attribute annotation turned out to be empty? If yes, can you provide us with a list of image ids where attributes were annotated, or some other way to find them?
", "votes": null, "replies": [ { "id": 528053, "author_name": "makeitworkjml", "author_url": "", "post_date": "05/07/2019 01:27:56", "content": "hi @lopuhin, we made sure that for images that do have attributes, only category id from 0 - 12 (main apparel categories) have attributes. \nImage with ClassId from \"0\" to \"12\" means that this image did not get annotated with fine-grained attributes.
Hope this helps!
", "votes": null, "replies": [] }, { "id": 528348, "author_name": "lopuhin", "author_url": "", "post_date": "05/07/2019 14:44:40", "content": "Aha, thanks for clarification @makeitworkjml this makes sense.\nFor anyone who might think that two sentences are contradictory, like me at first, pay attention that category is not the same as ClassId, and ClassId \"0\" means that not attributes are present :)
", "votes": null, "replies": [] }, { "id": 529499, "author_name": "clannad", "author_url": "", "post_date": "05/10/2019 03:42:40", "content": "hi @makeitworkjml , sorry I'm still a little bit confused. According to my statistics, there are also some categories annotated with attributes labels. If I don't make mistakes, class 27, 28 and 33 all contain one image annotated with attributes. Plus, I'm also curious whether there exists some category specific attribute. For example, there are only 14 instances belonging to cat12, where only 23 kinds of attributes appear in the cat12 training data. Should we assume that only these 23 attributes are used to describe cat12? Or other attributes can also be predicted for cat12 instances?
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "523116": "It seems like most of the examples have only one class_id. Is this class_id a category label? \n\nIf the example has more than one class_id (ie. '10_3_20_33_60_70_91'), do we assume the first one is a category label and the rest are attributes? \n\nIt's a little confusing because the category and attribute labels overlap (they both start at index 0).", "523135": "You are right, most of the training images (~80%) have one `ClassId`, which represents the main apparel category, like \"jacket\", or \"sleeve\", or \"glasses\". Out of those images that have attributes, only categories that are \"main apparel categories\"(category id from 0 - 12), like jacket, dress, pants, have attributes.\n\nThe first integer *is* a category label. And from second integer onwards, all labels are attributes. Both category and attribute are individually indexed.", "523146": "Thanks!", "525084": "makeitworkjml a related question, we should interpret those images that don't have attributes as images where attributes are not labelled, but could potentially be present, right? So in our ideal test submission we should have predicted attributes for such images (if they were in test)?", "525231": "lopuhin you are correct. In the test submission, all the masks in the test set were fully labeled.", "526271": "According to the description, the attributes are concatenated with the class ID. How exactly is this done? I was thinking if a particular image has a category and its attributes, its classID would be labelled as 'category, attribute1, attribute2' but that doesn't seem to be the case.", "526785": "Hi Sidhant, in this competition, we concatenated with \"_\", instead of \", \". But the basic idea is the same. \"category1_attribute12_attribute30\" means this masks has category 1, and attributes 12 and 30.", "527940": "makeitworkjml I have a related question - how to determine if attributes are labelled or not in the train set? Can we assume that if there is at least one attribute present in an image, then all objects in this image have attributes annotated, even if they are empty for some objects? Can there be objects in the training set that were annotated for attributes, and resulting attribute annotation turned out to be empty? If yes, can you provide us with a list of image ids where attributes were annotated, or some other way to find them?", "528053": "hi @lopuhin, we made sure that for images that do have attributes, only category id from 0 - 12 (main apparel categories) have attributes. \nImage with `ClassId` from \"0\" to \"12\" means that this image did not get annotated with fine-grained attributes.\n\nHope this helps!", "528348": "Aha, thanks for clarification @makeitworkjml this makes sense.\nFor anyone who might think that two sentences are contradictory, like me at first, pay attention that category is not the same as ClassId, and ClassId \"0\" means that not attributes are present :)", "528706": "Hi Menglin,\ncould you comment regarding the evaluation with the attributes? Do we have to get a category and all attributes right in order to get a count for the segmented object? (to be clear: if I miss one attribute, but all the rest is correct and IoU is 0.8 it will not count at all)", "529499": "hi @makeitworkjml , sorry I'm still a little bit confused. According to my statistics, there are also some categories annotated with attributes labels. If I don't make mistakes, class 27, 28 and 33 all contain one image annotated with attributes. Plus, I'm also curious whether there exists some category specific attribute. For example, there are only 14 instances belonging to cat12, where only 23 kinds of attributes appear in the cat12 training data. Should we assume that only these 23 attributes are used to describe cat12? Or other attributes can also be predicted for cat12 instances?" }, "source": "meta" }