Hello everyone, I had a query regarding how one could incorporate images with no wheat heads within training i.e. how do we represent it in the input dataframe ?
\n\nAs per EDA, in the dataset provided there are 49 images within training directory with no wheatheads detected. Currently I am taking images with >=1 bounding boxes i.e. as per what is present in train.csv.
\n\nPlease refer to my notebook below:\nhttps://www.kaggle.com/shebinscaria/pytorch-resnet50fpn-gwd
", "messages": [ { "id": "853290", "postDate": "05/19/2020 04:33:45", "content": "Hello everyone, I had a query regarding how one could incorporate images with no wheat heads within training i.e. how do we represent it in the input dataframe ?
\n\nAs per EDA, in the dataset provided there are 49 images within training directory with no wheatheads detected. Currently I am taking images with >=1 bounding boxes i.e. as per what is present in train.csv.
\n\nPlease refer to my notebook below:\nhttps://www.kaggle.com/shebinscaria/pytorch-resnet50fpn-gwd
", "rawMarkdown": "Hello everyone, I had a query regarding how one could incorporate images with no wheat heads within training i.e. how do we represent it in the input dataframe ? \n\nAs per EDA, in the dataset provided there are 49 images within training directory with no wheatheads detected. Currently I am taking images with >=1 bounding boxes i.e. as per what is present in train.csv.\n\nPlease refer to my notebook below:\nhttps://www.kaggle.com/shebinscaria/pytorch-resnet50fpn-gwd", "votes": null }, { "id": "854365", "postDate": "05/20/2020 01:52:42", "content": "Cant we label the image as background class ? Not sure what would be the bounding box co-ordinates though
", "rawMarkdown": "Cant we label the image as background class ? Not sure what would be the bounding box co-ordinates though", "votes": null }, { "id": "929016", "postDate": "07/14/2020 11:59:15", "content": "Can we annotate those 49 images as 1 0 0 0 0 in yolo format where 1 because they don't contatin wheat heads and 0 as they have no cordinates for wheat head . whereas 0 _ _ _ _ , where _ _ _ _ refers to bbox for all the 3373 images in train folder?
", "rawMarkdown": "Can we annotate those 49 images as **1 0 0 0 0** in yolo format where **1** because they don't contatin wheat heads and 0 as they have no cordinates for wheat head . whereas **0 _ _ _ _** , where _ _ _ _ refers to bbox for all the 3373 images in train folder?", "votes": null }, { "id": "929629", "postDate": "07/14/2020 20:00:48", "content": "It will depend on what framework you are using. In torchvision version 0.6.1, you can simply pass empty tensors like so\n\ntarget[\"boxes\"] = torch.zeros((0, 4), dtype=torch.float32)\ntarget[\"labels\"] = torch.zeros(0, dtype=torch.int64)\n
And for your submission.csv, make sure your PredictionString is just an empty string when no boxes are detected instead of a NaN value.
", "rawMarkdown": "It will depend on what framework you are using. In torchvision version 0.6.1, you can simply pass empty tensors like so\n```\ntarget[\"boxes\"] = torch.zeros((0, 4), dtype=torch.float32)\ntarget[\"labels\"] = torch.zeros(0, dtype=torch.int64)\n```\n\nAnd for your submission.csv, make sure your PredictionString is just an empty string when no boxes are detected instead of a NaN value.", "votes": null } ], "comments": [ { "id": 854365, "author_name": "kuramakyubi", "author_url": "", "post_date": "05/20/2020 01:52:42", "content": "Cant we label the image as background class ? Not sure what would be the bounding box co-ordinates though
", "votes": null, "replies": [] }, { "id": 929016, "author_name": "soumochatterjee", "author_url": "", "post_date": "07/14/2020 11:59:15", "content": "Can we annotate those 49 images as 1 0 0 0 0 in yolo format where 1 because they don't contatin wheat heads and 0 as they have no cordinates for wheat head . whereas 0 _ _ _ _ , where _ _ _ _ refers to bbox for all the 3373 images in train folder?
", "votes": null, "replies": [ { "id": 929629, "author_name": "matthewmasters", "author_url": "", "post_date": "07/14/2020 20:00:48", "content": "It will depend on what framework you are using. In torchvision version 0.6.1, you can simply pass empty tensors like so\n\ntarget[\"boxes\"] = torch.zeros((0, 4), dtype=torch.float32)\ntarget[\"labels\"] = torch.zeros(0, dtype=torch.int64)\n
And for your submission.csv, make sure your PredictionString is just an empty string when no boxes are detected instead of a NaN value.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "853290": "Hello everyone, I had a query regarding how one could incorporate images with no wheat heads within training i.e. how do we represent it in the input dataframe ? \n\nAs per EDA, in the dataset provided there are 49 images within training directory with no wheatheads detected. Currently I am taking images with >=1 bounding boxes i.e. as per what is present in train.csv.\n\nPlease refer to my notebook below:\nhttps://www.kaggle.com/shebinscaria/pytorch-resnet50fpn-gwd", "854365": "Cant we label the image as background class ? Not sure what would be the bounding box co-ordinates though", "929016": "Can we annotate those 49 images as **1 0 0 0 0** in yolo format where **1** because they don't contatin wheat heads and 0 as they have no cordinates for wheat head . whereas **0 _ _ _ _** , where _ _ _ _ refers to bbox for all the 3373 images in train folder?", "929629": "It will depend on what framework you are using. In torchvision version 0.6.1, you can simply pass empty tensors like so\n```\ntarget[\"boxes\"] = torch.zeros((0, 4), dtype=torch.float32)\ntarget[\"labels\"] = torch.zeros(0, dtype=torch.int64)\n```\n\nAnd for your submission.csv, make sure your PredictionString is just an empty string when no boxes are detected instead of a NaN value." }, "source": "meta" }