Looking at the data, there are more than 20Gb of images. Grid_sizes has 450 unique imageIDs, however train_wkt has only 22. Why do we have all of these images in the dataset if we only have labelled training data for less than 5%?
\n\nAm I missing something here?
", "messages": [ { "id": "150844", "postDate": "12/17/2016 01:29:49", "content": "Looking at the data, there are more than 20Gb of images. Grid_sizes has 450 unique imageIDs, however train_wkt has only 22. Why do we have all of these images in the dataset if we only have labelled training data for less than 5%?
\n\nAm I missing something here?
", "rawMarkdown": "Looking at the data, there are more than 20Gb of images. Grid_sizes has 450 unique imageIDs, however train_wkt has only 22. Why do we have all of these images in the dataset if we only have labelled training data for less than 5%?\r\n\r\nAm I missing something here?", "votes": null }, { "id": "150860", "postDate": "12/17/2016 03:58:53", "content": "The remaining images are the test set:
\n\nEverything seems to sum up correctly.
", "rawMarkdown": "The remaining images are the test set:\r\n\r\n - sample_submission.csv has the 429 unique ImageIDs.\r\n - train_wkt.csv has 21 unique ImageIDs (you incorrectly said 22).\r\n - grid_sizes.csv has 450 ImageIDs.\r\n - three_band/ has total 450 images.\r\n - sixteen_band/ has 1350 images (3 images for each of the 450 ImageIDs = 1350 ).\r\n\r\nEverything seems to sum up correctly.", "votes": null }, { "id": "150862", "postDate": "12/17/2016 04:04:35", "content": "No, the updated v2 train_wkt.csv has 22 imageIDs, check the data page again for the new file.
\n\nThe only thing that sums up is that 3*450 = 1350. It doesn't change the fact that there are only labels for 22 images, the remaining 428 are unlabelled images which makes them a part of the test set. It just seems odd to me that the test set is more than 20 times larger than the training set.
\n\nI guess my question was: does this seem strange to anyone else? Or am I missing something crucial here?
", "rawMarkdown": "No, the updated v2 train_wkt.csv has 22 imageIDs, check the data page again for the new file.\r\n\r\nThe only thing that sums up is that 3*450 = 1350. It doesn't change the fact that there are only labels for 22 images, the remaining 428 are unlabelled images which makes them a part of the test set. It just seems odd to me that the test set is more than 20 times larger than the training set.\r\n\r\nI guess my question was: does this seem strange to anyone else? Or am I missing something crucial here?", "votes": null }, { "id": "150863", "postDate": "12/17/2016 04:07:21", "content": "\n\n\ntrain_wkt.csv has 21 unique ImageIDs (you incorrectly said 22).
\n
You were correct. It was announced in another thread that another image was added to the train set -- the train_wkt_v2.csv does indeed have 22 ImageIDs. Sorry for the confusion.
", "rawMarkdown": "> train_wkt.csv has 21 unique ImageIDs (you incorrectly said 22).\r\n\r\nYou were correct. It was announced in another thread that another image was added to the train set -- the train_wkt_v2.csv does indeed have 22 ImageIDs. Sorry for the confusion.", "votes": null }, { "id": "150954", "postDate": "12/17/2016 20:27:27", "content": "[quote=grantbey;150862]
\n\nI guess my question was: does this seem strange to anyone else? Or am I missing something crucial here?
\n\n[/quote]
\n\n@grantbey, not strange at all that there's a large test set. It makes it harder for any hand labeling during the competition and final judgement. It's also more real-worldish as in practice one would want to develop a good model and then apply it everywhere to save the most human resources.
", "rawMarkdown": "[quote=grantbey;150862]\r\n\r\nI guess my question was: does this seem strange to anyone else? Or am I missing something crucial here?\r\n\r\n[/quote]\r\n\r\n@grantbey, not strange at all that there's a large test set. It makes it harder for any hand labeling during the competition and final judgement. It's also more real-worldish as in practice one would want to develop a good model and then apply it *everywhere* to save the most human resources.", "votes": null }, { "id": "150956", "postDate": "12/17/2016 20:46:00", "content": "[quote=zero zero;150954]
\n\n@grantbey, not strange at all that there's a large test set. It makes it harder for any hand labeling during the competition and final judgement. It's also more real-worldish as in practice one would want to develop a good model and then apply it everywhere to save the most human resources.
\n\n[/quote]
\n\nThanks, I guess I just thought it was unusual given other ML applications I've played with. I was mostly confirming that what I was seeing was correct and that I wasn't missing some crucial point and thus ignoring tons of training data.
\n\nThanks for your response!
", "rawMarkdown": "[quote=zero zero;150954]\r\n\r\n@grantbey, not strange at all that there's a large test set. It makes it harder for any hand labeling during the competition and final judgement. It's also more real-worldish as in practice one would want to develop a good model and then apply it *everywhere* to save the most human resources.\r\n\r\n[/quote]\r\n\r\nThanks, I guess I just thought it was unusual given other ML applications I've played with. I was mostly confirming that what I was seeing was correct and that I wasn't missing some crucial point and thus ignoring tons of training data.\r\n\r\nThanks for your response!", "votes": null }, { "id": "387758", "postDate": "09/15/2018 16:06:27", "content": "Manually tagging the data is not more a hurdle. Check this out,
\n\nhttps://www.youtube.com/watch?v=tYqnsp-OcLQ
\n\nhttps://www.youtube.com/watch?v=xkSEnDIlvhI
\n\nyou can quickly create your own image and video segmentation data in no time!!
\n\nNeurala’s Brain Builder Tool: https://hubs.ly/H0dkvRK0
", "rawMarkdown": "Manually tagging the data is not more a hurdle. Check this out,\n\nhttps://www.youtube.com/watch?v=tYqnsp-OcLQ\n\nhttps://www.youtube.com/watch?v=xkSEnDIlvhI\n\nyou can quickly create your own image and video segmentation data in no time!!\n\nNeurala’s Brain Builder Tool: https://hubs.ly/H0dkvRK0", "votes": null } ], "comments": [ { "id": 150860, "author_name": "amanbh", "author_url": "", "post_date": "12/17/2016 03:58:53", "content": "The remaining images are the test set:
\n\nEverything seems to sum up correctly.
", "votes": null, "replies": [] }, { "id": 150862, "author_name": "grantbey", "author_url": "", "post_date": "12/17/2016 04:04:35", "content": "No, the updated v2 train_wkt.csv has 22 imageIDs, check the data page again for the new file.
\n\nThe only thing that sums up is that 3*450 = 1350. It doesn't change the fact that there are only labels for 22 images, the remaining 428 are unlabelled images which makes them a part of the test set. It just seems odd to me that the test set is more than 20 times larger than the training set.
\n\nI guess my question was: does this seem strange to anyone else? Or am I missing something crucial here?
", "votes": null, "replies": [ { "id": 150954, "author_name": "zerozero", "author_url": "", "post_date": "12/17/2016 20:27:27", "content": "[quote=grantbey;150862]
\n\nI guess my question was: does this seem strange to anyone else? Or am I missing something crucial here?
\n\n[/quote]
\n\n@grantbey, not strange at all that there's a large test set. It makes it harder for any hand labeling during the competition and final judgement. It's also more real-worldish as in practice one would want to develop a good model and then apply it everywhere to save the most human resources.
", "votes": null, "replies": [ { "id": 150956, "author_name": "grantbey", "author_url": "", "post_date": "12/17/2016 20:46:00", "content": "[quote=zero zero;150954]
\n\n@grantbey, not strange at all that there's a large test set. It makes it harder for any hand labeling during the competition and final judgement. It's also more real-worldish as in practice one would want to develop a good model and then apply it everywhere to save the most human resources.
\n\n[/quote]
\n\nThanks, I guess I just thought it was unusual given other ML applications I've played with. I was mostly confirming that what I was seeing was correct and that I wasn't missing some crucial point and thus ignoring tons of training data.
\n\nThanks for your response!
", "votes": null, "replies": [] } ] }, { "id": 387758, "author_name": "ajagetia", "author_url": "", "post_date": "09/15/2018 16:06:27", "content": "Manually tagging the data is not more a hurdle. Check this out,
\n\nhttps://www.youtube.com/watch?v=tYqnsp-OcLQ
\n\nhttps://www.youtube.com/watch?v=xkSEnDIlvhI
\n\nyou can quickly create your own image and video segmentation data in no time!!
\n\nNeurala’s Brain Builder Tool: https://hubs.ly/H0dkvRK0
", "votes": null, "replies": [] } ] }, { "id": 150863, "author_name": "amanbh", "author_url": "", "post_date": "12/17/2016 04:07:21", "content": "\n\n\ntrain_wkt.csv has 21 unique ImageIDs (you incorrectly said 22).
\n
You were correct. It was announced in another thread that another image was added to the train set -- the train_wkt_v2.csv does indeed have 22 ImageIDs. Sorry for the confusion.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "150844": "Looking at the data, there are more than 20Gb of images. Grid_sizes has 450 unique imageIDs, however train_wkt has only 22. Why do we have all of these images in the dataset if we only have labelled training data for less than 5%?\r\n\r\nAm I missing something here?", "150860": "The remaining images are the test set:\r\n\r\n - sample_submission.csv has the 429 unique ImageIDs.\r\n - train_wkt.csv has 21 unique ImageIDs (you incorrectly said 22).\r\n - grid_sizes.csv has 450 ImageIDs.\r\n - three_band/ has total 450 images.\r\n - sixteen_band/ has 1350 images (3 images for each of the 450 ImageIDs = 1350 ).\r\n\r\nEverything seems to sum up correctly.", "150862": "No, the updated v2 train_wkt.csv has 22 imageIDs, check the data page again for the new file.\r\n\r\nThe only thing that sums up is that 3*450 = 1350. It doesn't change the fact that there are only labels for 22 images, the remaining 428 are unlabelled images which makes them a part of the test set. It just seems odd to me that the test set is more than 20 times larger than the training set.\r\n\r\nI guess my question was: does this seem strange to anyone else? Or am I missing something crucial here?", "150863": "> train_wkt.csv has 21 unique ImageIDs (you incorrectly said 22).\r\n\r\nYou were correct. It was announced in another thread that another image was added to the train set -- the train_wkt_v2.csv does indeed have 22 ImageIDs. Sorry for the confusion.", "150954": "[quote=grantbey;150862]\r\n\r\nI guess my question was: does this seem strange to anyone else? Or am I missing something crucial here?\r\n\r\n[/quote]\r\n\r\n@grantbey, not strange at all that there's a large test set. It makes it harder for any hand labeling during the competition and final judgement. It's also more real-worldish as in practice one would want to develop a good model and then apply it *everywhere* to save the most human resources.", "150956": "[quote=zero zero;150954]\r\n\r\n@grantbey, not strange at all that there's a large test set. It makes it harder for any hand labeling during the competition and final judgement. It's also more real-worldish as in practice one would want to develop a good model and then apply it *everywhere* to save the most human resources.\r\n\r\n[/quote]\r\n\r\nThanks, I guess I just thought it was unusual given other ML applications I've played with. I was mostly confirming that what I was seeing was correct and that I wasn't missing some crucial point and thus ignoring tons of training data.\r\n\r\nThanks for your response!", "387758": "Manually tagging the data is not more a hurdle. Check this out,\n\nhttps://www.youtube.com/watch?v=tYqnsp-OcLQ\n\nhttps://www.youtube.com/watch?v=xkSEnDIlvhI\n\nyou can quickly create your own image and video segmentation data in no time!!\n\nNeurala’s Brain Builder Tool: https://hubs.ly/H0dkvRK0" }, "source": "meta" }