![\"enter](\"http://image-net.org/challenges/LSVRC/2014/ILSVRC2012_val_00004465.png\" "\\"\\"")
Could you kindly state whether each pixel has a definite class label or not? Is it possible that one pixel has multiple class labels?
\n\nI am asking because it seems to me that the polygons of different classes intersect each other. If they do intersect and each pixel only has a single class label, then we need a fixed order of projection in order to uniquely determine the segmentation map.
\n\nAttached you find an animation that shows the intersection of polygons of different classes.
", "messages": [ { "id": "150576", "postDate": "12/15/2016 20:53:20", "content": "Could you kindly state whether each pixel has a definite class label or not? Is it possible that one pixel has multiple class labels?
\n\nI am asking because it seems to me that the polygons of different classes intersect each other. If they do intersect and each pixel only has a single class label, then we need a fixed order of projection in order to uniquely determine the segmentation map.
\n\nAttached you find an animation that shows the intersection of polygons of different classes.
", "rawMarkdown": "Could you kindly state whether each pixel has a definite class label or not? Is it possible that one pixel has multiple class labels?\r\n\r\nI am asking because it seems to me that the polygons of different classes intersect each other. If they do intersect and each pixel only has a single class label, then we need a fixed order of projection in order to uniquely determine the segmentation map. \r\n\r\nAttached you find an animation that shows the intersection of polygons of different classes.", "votes": null }, { "id": "150770", "postDate": "12/16/2016 16:56:53", "content": "I agree that an answer to this would be very useful. As you point out, if using a segmentation approach one would want each pixel to have an exclusive label. However, if we do not get a reply from the administrator, I think we should proceed under the assumption of multi-label....ie, a pixel could be a road and a vehicle, etc..
\n\nI have been wondering why we must submit in polygon form when a multi-band compressed .tif file would suffice and be easier.
", "rawMarkdown": "I agree that an answer to this would be very useful. As you point out, if using a segmentation approach one would want each pixel to have an exclusive label. However, if we do not get a reply from the administrator, I think we should proceed under the assumption of multi-label....ie, a pixel could be a road *and* a vehicle, etc.. \r\n\r\nI have been wondering why we must submit in polygon form when a multi-band compressed .tif file would suffice and be easier.", "votes": null }, { "id": "150785", "postDate": "12/16/2016 18:17:26", "content": "The labels do sometimes overlap. For example, a house may be right next to a road, and these two polygons might intersect slightly. Another example is a vehicle is parked on the road, then those two will overlap.
\n\nAs for the submission format: it was a request from Dstl (the competition host) that the submission be vector-based and not pixel-based such as a tiff file. Vector-based output is the norm in the geospatial domain, so we decided to make the input/output of this competition useful for them.
", "rawMarkdown": "The labels do *sometimes* overlap. For example, a house may be right next to a road, and these two polygons might intersect slightly. Another example is a vehicle is parked on the road, then those two will overlap. \r\n\r\nAs for the submission format: it was a request from Dstl (the competition host) that the submission be vector-based and not pixel-based such as a tiff file. Vector-based output is the norm in the geospatial domain, so we decided to make the input/output of this competition useful for them.", "votes": null }, { "id": "150786", "postDate": "12/16/2016 18:20:04", "content": "My guess is that you have to predict the presence of all classes on an image. For example, for image 6010_0_0 your submission file will look like this:
\n\n6010_0_0, 1, MULTIPOLYGON ()
\n\n6010_0_0, 2, MULTIPOLYGON ()
\n\n...
\n\n6010_0_0, 10, MULTIPOLYGON ()
\n\nIf you didn't found an object of class N on that image, then you set it to be MULTIPOLYGON EMPTY.
", "rawMarkdown": "My guess is that you have to predict the presence of all classes on an image. For example, for image 6010_0_0 your submission file will look like this:\r\n\r\n6010_0_0, 1, MULTIPOLYGON ()\r\n\r\n6010_0_0, 2, MULTIPOLYGON ()\r\n\r\n...\r\n\r\n6010_0_0, 10, MULTIPOLYGON ()\r\n\r\nIf you didn't found an object of class N on that image, then you set it to be MULTIPOLYGON EMPTY.", "votes": null }, { "id": "150835", "postDate": "12/16/2016 23:29:49", "content": "[quote=Wendy Kan;150785]
\n\nThe labels do sometimes overlap. For example, a house may be right next to a road, and these two polygons might intersect slightly. Another example is a vehicle is parked on the road, then those two will overlap.
\n\nAs for the submission format: it was a request from Dstl (the competition host) that the submission be vector-based and not pixel-based such as a tiff file. Vector-based output is the norm in the geospatial domain, so we decided to make the input/output of this competition useful for them.
\n\n[/quote]
\n\nHi Wendy, thanks for some clarification. However sometimes is a bit vague! Do you mean that in the small number of cases where polygons intersect, we should just pick a class? Or do you mean that a pixel can be assigned to 2(or more) different classes. So far, I am understanding that this is a problem where pixels can have more than one label...which is perfectly fine; but a committed confirmation would be nice.
\n\nAnd, yes, many things in the GIS world are polygon based...however most processing and classifications are done with rasters and certain criteria are established prior to converting into polygons...things like the minimum mapping unit and any common filtering.
\n\nAlso, prior to processing/modeling, most of us will be constructing truth masks in raster form and to do this it would be nice to know the client's rasterization rules. ie. a pixel belongs to a polygon if: 1) the polygon contains the center of the pixel(usual method) or 2) the line(drawn by the polygon) touches any part of the square pixel. It may not seem like a big deal but for classes with many small polygons, it could make a difference. (cross-validating both strategies and picking the better one is not the correct way to find this out!) Seeing an example of something in raster form(from the client) would quickly answer this as we could compare it to one of their polygons...like a one band raster for a single class for a single training image.
", "rawMarkdown": "[quote=Wendy Kan;150785]\r\n\r\nThe labels do *sometimes* overlap. For example, a house may be right next to a road, and these two polygons might intersect slightly. Another example is a vehicle is parked on the road, then those two will overlap. \r\n\r\nAs for the submission format: it was a request from Dstl (the competition host) that the submission be vector-based and not pixel-based such as a tiff file. Vector-based output is the norm in the geospatial domain, so we decided to make the input/output of this competition useful for them. \r\n\r\n[/quote]\r\n\r\nHi Wendy, thanks for some clarification. However *sometimes* is a bit vague! Do you mean that in the small number of cases where polygons intersect, we should just pick a class? Or do you mean that a pixel can be assigned to 2(or more) different classes. So far, I am understanding that this is a problem where pixels can have more than one label...which is perfectly fine; but a committed confirmation would be nice.\r\n\r\nAnd, yes, many things in the GIS world are polygon based...however most processing and classifications are done with rasters and certain criteria are established prior to converting into polygons...things like the minimum mapping unit and any common filtering. \r\n\r\nAlso, prior to processing/modeling, most of us will be constructing truth masks in raster form and to do this it would be nice to know the client's rasterization rules. ie. a pixel belongs to a polygon if: 1) the polygon contains the center of the pixel(usual method) or 2) the line(drawn by the polygon) touches any part of the square pixel. It may not seem like a big deal but for classes with many small polygons, it could make a difference. (cross-validating both strategies and picking the better one is *not* the correct way to find this out!) Seeing an example of something in raster form(from the client) would quickly answer this as we could compare it to one of their polygons...like a one band raster for a single class for a single training image.", "votes": null }, { "id": "150935", "postDate": "12/17/2016 17:59:37", "content": "@zerozero,
\n\nI meant the latter. One pixel may belong to multiple labels. An easier example from ImageNet:
\n\n
as you see the object in the image is overlapped. What I meant by they sometimes overlap is that these are satellite images, so objects don't overlap very much, as in the above image they overlap a lot.
\n\nI'll check on the labeling guidelines that Dstl used.
", "rawMarkdown": "zerozero, \r\n\r\nI meant the latter. One pixel may belong to multiple labels. An easier example from ImageNet: \r\n\r\n![enter image description here][1]\r\n\r\nas you see the object in the image is overlapped. What I meant by they _sometimes_ overlap is that these are satellite images, so objects don't overlap _very much_, as in the above image they overlap _a lot_. \r\n\r\nI'll check on the labeling guidelines that Dstl used. \r\n\r\n\r\n [1]: http://image-net.org/challenges/LSVRC/2014/ILSVRC2012_val_00004465.png", "votes": null } ], "comments": [ { "id": 150770, "author_name": "zerozero", "author_url": "", "post_date": "12/16/2016 16:56:53", "content": "I agree that an answer to this would be very useful. As you point out, if using a segmentation approach one would want each pixel to have an exclusive label. However, if we do not get a reply from the administrator, I think we should proceed under the assumption of multi-label....ie, a pixel could be a road and a vehicle, etc..
\n\nI have been wondering why we must submit in polygon form when a multi-band compressed .tif file would suffice and be easier.
", "votes": null, "replies": [] }, { "id": 150785, "author_name": "wendykan", "author_url": "", "post_date": "12/16/2016 18:17:26", "content": "The labels do sometimes overlap. For example, a house may be right next to a road, and these two polygons might intersect slightly. Another example is a vehicle is parked on the road, then those two will overlap.
\n\nAs for the submission format: it was a request from Dstl (the competition host) that the submission be vector-based and not pixel-based such as a tiff file. Vector-based output is the norm in the geospatial domain, so we decided to make the input/output of this competition useful for them.
", "votes": null, "replies": [ { "id": 150835, "author_name": "zerozero", "author_url": "", "post_date": "12/16/2016 23:29:49", "content": "[quote=Wendy Kan;150785]
\n\nThe labels do sometimes overlap. For example, a house may be right next to a road, and these two polygons might intersect slightly. Another example is a vehicle is parked on the road, then those two will overlap.
\n\nAs for the submission format: it was a request from Dstl (the competition host) that the submission be vector-based and not pixel-based such as a tiff file. Vector-based output is the norm in the geospatial domain, so we decided to make the input/output of this competition useful for them.
\n\n[/quote]
\n\nHi Wendy, thanks for some clarification. However sometimes is a bit vague! Do you mean that in the small number of cases where polygons intersect, we should just pick a class? Or do you mean that a pixel can be assigned to 2(or more) different classes. So far, I am understanding that this is a problem where pixels can have more than one label...which is perfectly fine; but a committed confirmation would be nice.
\n\nAnd, yes, many things in the GIS world are polygon based...however most processing and classifications are done with rasters and certain criteria are established prior to converting into polygons...things like the minimum mapping unit and any common filtering.
\n\nAlso, prior to processing/modeling, most of us will be constructing truth masks in raster form and to do this it would be nice to know the client's rasterization rules. ie. a pixel belongs to a polygon if: 1) the polygon contains the center of the pixel(usual method) or 2) the line(drawn by the polygon) touches any part of the square pixel. It may not seem like a big deal but for classes with many small polygons, it could make a difference. (cross-validating both strategies and picking the better one is not the correct way to find this out!) Seeing an example of something in raster form(from the client) would quickly answer this as we could compare it to one of their polygons...like a one band raster for a single class for a single training image.
", "votes": null, "replies": [] } ] }, { "id": 150786, "author_name": "alexlu", "author_url": "", "post_date": "12/16/2016 18:20:04", "content": "My guess is that you have to predict the presence of all classes on an image. For example, for image 6010_0_0 your submission file will look like this:
\n\n6010_0_0, 1, MULTIPOLYGON ()
\n\n6010_0_0, 2, MULTIPOLYGON ()
\n\n...
\n\n6010_0_0, 10, MULTIPOLYGON ()
\n\nIf you didn't found an object of class N on that image, then you set it to be MULTIPOLYGON EMPTY.
", "votes": null, "replies": [] }, { "id": 150935, "author_name": "wendykan", "author_url": "", "post_date": "12/17/2016 17:59:37", "content": "@zerozero,
\n\nI meant the latter. One pixel may belong to multiple labels. An easier example from ImageNet:
\n\n
as you see the object in the image is overlapped. What I meant by they sometimes overlap is that these are satellite images, so objects don't overlap very much, as in the above image they overlap a lot.
\n\nI'll check on the labeling guidelines that Dstl used.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "150576": "Could you kindly state whether each pixel has a definite class label or not? Is it possible that one pixel has multiple class labels?\r\n\r\nI am asking because it seems to me that the polygons of different classes intersect each other. If they do intersect and each pixel only has a single class label, then we need a fixed order of projection in order to uniquely determine the segmentation map. \r\n\r\nAttached you find an animation that shows the intersection of polygons of different classes.", "150770": "I agree that an answer to this would be very useful. As you point out, if using a segmentation approach one would want each pixel to have an exclusive label. However, if we do not get a reply from the administrator, I think we should proceed under the assumption of multi-label....ie, a pixel could be a road *and* a vehicle, etc.. \r\n\r\nI have been wondering why we must submit in polygon form when a multi-band compressed .tif file would suffice and be easier.", "150785": "The labels do *sometimes* overlap. For example, a house may be right next to a road, and these two polygons might intersect slightly. Another example is a vehicle is parked on the road, then those two will overlap. \r\n\r\nAs for the submission format: it was a request from Dstl (the competition host) that the submission be vector-based and not pixel-based such as a tiff file. Vector-based output is the norm in the geospatial domain, so we decided to make the input/output of this competition useful for them.", "150786": "My guess is that you have to predict the presence of all classes on an image. For example, for image 6010_0_0 your submission file will look like this:\r\n\r\n6010_0_0, 1, MULTIPOLYGON ()\r\n\r\n6010_0_0, 2, MULTIPOLYGON ()\r\n\r\n...\r\n\r\n6010_0_0, 10, MULTIPOLYGON ()\r\n\r\nIf you didn't found an object of class N on that image, then you set it to be MULTIPOLYGON EMPTY.", "150835": "[quote=Wendy Kan;150785]\r\n\r\nThe labels do *sometimes* overlap. For example, a house may be right next to a road, and these two polygons might intersect slightly. Another example is a vehicle is parked on the road, then those two will overlap. \r\n\r\nAs for the submission format: it was a request from Dstl (the competition host) that the submission be vector-based and not pixel-based such as a tiff file. Vector-based output is the norm in the geospatial domain, so we decided to make the input/output of this competition useful for them. \r\n\r\n[/quote]\r\n\r\nHi Wendy, thanks for some clarification. However *sometimes* is a bit vague! Do you mean that in the small number of cases where polygons intersect, we should just pick a class? Or do you mean that a pixel can be assigned to 2(or more) different classes. So far, I am understanding that this is a problem where pixels can have more than one label...which is perfectly fine; but a committed confirmation would be nice.\r\n\r\nAnd, yes, many things in the GIS world are polygon based...however most processing and classifications are done with rasters and certain criteria are established prior to converting into polygons...things like the minimum mapping unit and any common filtering. \r\n\r\nAlso, prior to processing/modeling, most of us will be constructing truth masks in raster form and to do this it would be nice to know the client's rasterization rules. ie. a pixel belongs to a polygon if: 1) the polygon contains the center of the pixel(usual method) or 2) the line(drawn by the polygon) touches any part of the square pixel. It may not seem like a big deal but for classes with many small polygons, it could make a difference. (cross-validating both strategies and picking the better one is *not* the correct way to find this out!) Seeing an example of something in raster form(from the client) would quickly answer this as we could compare it to one of their polygons...like a one band raster for a single class for a single training image.", "150935": "zerozero, \r\n\r\nI meant the latter. One pixel may belong to multiple labels. An easier example from ImageNet: \r\n\r\n![enter image description here][1]\r\n\r\nas you see the object in the image is overlapped. What I meant by they _sometimes_ overlap is that these are satellite images, so objects don't overlap _very much_, as in the above image they overlap _a lot_. \r\n\r\nI'll check on the labeling guidelines that Dstl used. \r\n\r\n\r\n [1]: http://image-net.org/challenges/LSVRC/2014/ILSVRC2012_val_00004465.png" }, "source": "meta" }