The competition provides a \"detection-bbox.csv\" and a \"human-imagelabels.csv\". But how to use the \"human-imagelabels.csv\"?
", "messages": [ { "id": 567593, "postDate": "2019-07-03T18:29:10.140Z", "content": "You need to use both of these in order to train your model. This data is sparsely labeled. Unlike datasets like COCO which in every image, EVERY class has bounding boxes annotated. human-imagelabels.csv has both positive and negative labels for each image. This means that if an image has positive labels [Person, Basketball] and negative labels [Tree, Apple], then EVERY instance of a Person and Basketball should have a bounding box associated with it (exception for GroupOf tagged objects), and EVERY instance of Tree and Apple is labeled. Since these are negative labels it's basically saying there are no trees and no apple objects in the image.
\n\nNormal training for Object Detection assumes Dense labeling, since this is sparse labeling, you cannot assume that if your model detects footware with probability of 0.7 in the image, it is incorrect (and thus increase your loss). Be careful about using of-the-shelf model training pipelines designed for other datasets like COCO! You will need to modify them to support sparse labeling.
\n\nSee the Open Image Dataset description for bounding boxes:
\n\n\n", "rawMarkdown": "You need to use both of these in order to train your model. This data is sparsely labeled. Unlike datasets like COCO which in every image, EVERY class has bounding boxes annotated. human-imagelabels.csv has both positive and negative labels for each image. This means that if an image has positive labels [Person, Basketball] and negative labels [Tree, Apple], then EVERY instance of a Person and Basketball should have a bounding box associated with it (exception for GroupOf tagged objects), and EVERY instance of Tree and Apple is labeled. Since these are negative labels it's basically saying there are no trees and no apple objects in the image.\n\nNormal training for Object Detection assumes Dense labeling, since this is sparse labeling, you cannot assume that if your model detects footware with probability of 0.7 in the image, it is incorrect (and thus increase your loss). Be careful about using of-the-shelf model training pipelines designed for other datasets like COCO! You will need to modify them to support sparse labeling.\n\nSee the Open Image Dataset description for bounding boxes:\n\n> For each label in an image, we exhaustively annotated every instance of that object class in the image (but see below for group cases).", "replies": [ { "id": 577536, "postDate": "2019-07-16T18:43:08.693Z", "content": "For each label in an image, we exhaustively annotated every instance of that object class in the image (but see below for group cases).
\n
\n\n\nFor the training set, we annotated boxes in 1.74M images, for the available positive human-verified image-level labels. We focused on the most specific labels. For example, if an image has labels {car, limousine, screwdriver}, we annotated boxes for limousine and screwdriver. For each label in an image, we exhaustively annotated every instance of that object class in the image (but see below for group cases).
\n
I took this to mean that every instance of the 600 available labels that are present in an image were annotated, not just the ones that got labeled in the human-imagelabels.csv. I am unsure though. The wording isn't super clear.
", "rawMarkdown": "> For the training set, we annotated boxes in 1.74M images, for the available positive human-verified image-level labels. We focused on the most specific labels. For example, if an image has labels {car, limousine, screwdriver}, we annotated boxes for limousine and screwdriver. For each label in an image, we exhaustively annotated every instance of that object class in the image (but see below for group cases). \n\nI took this to mean that every instance of the 600 available labels that are present in an image were annotated, not just the ones that got labeled in the human-imagelabels.csv. I am unsure though. The wording isn't super clear.", "isDeleted": true } ] }, { "id": 566692, "postDate": "2019-07-02T13:17:36.130Z", "content": "The competition provides a \"detection-bbox.csv\" and a \"human-imagelabels.csv\". But how to use the \"human-imagelabels.csv\"?
", "rawMarkdown": "The competition provides a \"detection-bbox.csv\" and a \"human-imagelabels.csv\". But how to use the \"human-imagelabels.csv\"? " }, { "id": 566707, "postDate": "2019-07-02T13:49:32.763Z", "content": "I know that many of the past competiton's teams mentioned that they could not find a way to include the image labels in way that helped their models.
", "rawMarkdown": "I know that many of the past competiton's teams mentioned that they could not find a way to include the image labels in way that helped their models. ", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 567593, "author_name": "Paul Johnson", "author_url": "", "post_date": "2019-07-03T18:29:10.140000", "content": "You need to use both of these in order to train your model. This data is sparsely labeled. Unlike datasets like COCO which in every image, EVERY class has bounding boxes annotated. human-imagelabels.csv has both positive and negative labels for each image. This means that if an image has positive labels [Person, Basketball] and negative labels [Tree, Apple], then EVERY instance of a Person and Basketball should have a bounding box associated with it (exception for GroupOf tagged objects), and EVERY instance of Tree and Apple is labeled. Since these are negative labels it's basically saying there are no trees and no apple objects in the image.
\n\nNormal training for Object Detection assumes Dense labeling, since this is sparse labeling, you cannot assume that if your model detects footware with probability of 0.7 in the image, it is incorrect (and thus increase your loss). Be careful about using of-the-shelf model training pipelines designed for other datasets like COCO! You will need to modify them to support sparse labeling.
\n\nSee the Open Image Dataset description for bounding boxes:
\n\n\n", "votes": 0, "replies": [ { "id": 577536, "author_name": "", "author_url": "", "post_date": "2019-07-16T18:43:08.693000", "content": "For each label in an image, we exhaustively annotated every instance of that object class in the image (but see below for group cases).
\n
\n\n\nFor the training set, we annotated boxes in 1.74M images, for the available positive human-verified image-level labels. We focused on the most specific labels. For example, if an image has labels {car, limousine, screwdriver}, we annotated boxes for limousine and screwdriver. For each label in an image, we exhaustively annotated every instance of that object class in the image (but see below for group cases).
\n
I took this to mean that every instance of the 600 available labels that are present in an image were annotated, not just the ones that got labeled in the human-imagelabels.csv. I am unsure though. The wording isn't super clear.
", "votes": 0, "replies": [] } ] }, { "id": 566707, "author_name": "", "author_url": "", "post_date": "2019-07-02T13:49:32.763000", "content": "I know that many of the past competiton's teams mentioned that they could not find a way to include the image labels in way that helped their models.
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "567593": "You need to use both of these in order to train your model. This data is sparsely labeled. Unlike datasets like COCO which in every image, EVERY class has bounding boxes annotated. human-imagelabels.csv has both positive and negative labels for each image. This means that if an image has positive labels [Person, Basketball] and negative labels [Tree, Apple], then EVERY instance of a Person and Basketball should have a bounding box associated with it (exception for GroupOf tagged objects), and EVERY instance of Tree and Apple is labeled. Since these are negative labels it's basically saying there are no trees and no apple objects in the image.\n\nNormal training for Object Detection assumes Dense labeling, since this is sparse labeling, you cannot assume that if your model detects footware with probability of 0.7 in the image, it is incorrect (and thus increase your loss). Be careful about using of-the-shelf model training pipelines designed for other datasets like COCO! You will need to modify them to support sparse labeling.\n\nSee the Open Image Dataset description for bounding boxes:\n\n> For each label in an image, we exhaustively annotated every instance of that object class in the image (but see below for group cases).", "566692": "The competition provides a \"detection-bbox.csv\" and a \"human-imagelabels.csv\". But how to use the \"human-imagelabels.csv\"? ", "566707": "I know that many of the past competiton's teams mentioned that they could not find a way to include the image labels in way that helped their models. " } }