![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1555050%2F4fb3adaeba5141bb44fc30a20cca7f5a%2F6.png?generation=1593610342732040&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1555050%2Fc0464500e497a0276eaca2429bdd06d0%2F1.png?generation=1593610342569905&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1555050%2Fc4131113d05f95fa15b112d8a131256d%2F4.png?generation=1593610342442139&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1555050%2F49afe21891958df194975088a157728e%2F5.png?generation=1593610340958000&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1555050%2Fd80c4e6cefff9f4b303dd928e8e2959e%2F3.png?generation=1593610337861238&alt=media\")
After quick error analysis I found multiple (hundreds) errors in train labels, some of them I attached. Also after removing them from train, I got high and stable cv score but lower lb score, so I suppose that test labels are also dirty.
\n\nCould you manual check test labelling and fix it if it's wrong? Or suggest another solution? @maggiemd @bendvd
\n\nI think that you want to get a good tool for labellling instead of overfitting to dirty labels.
\n\n\n\n\n\n
After quick error analysis I found multiple (hundreds) errors in train labels, some of them I attached. Also after removing them from train, I got high and stable cv score but lower lb score, so I suppose that test labels are also dirty.
\n\nCould you manual check test labelling and fix it if it's wrong? Or suggest another solution? @maggiemd @bendvd
\n\nI think that you want to get a good tool for labellling instead of overfitting to dirty labels.
\n\n\n\n\n\n
These come from the original images being split up into 1024 patches. I would hope that they at least fix these for the private test set, as they indicated in their last update. Though with the quality of the data being what it is... I wouldn't be too confident that it'll be get fixed properly.
", "rawMarkdown": "These come from the original images being split up into 1024 patches. I would hope that they at least fix these for the private test set, as they indicated in their last update. Though with the quality of the data being what it is... I wouldn't be too confident that it'll be get fixed properly.", "votes": null }, { "id": "912061", "postDate": "07/02/2020 07:23:42", "content": "I had a model which predicts a very tight bbox around the wheat-heads and it looked very promising but got LB score of lesser than .6, it was really saddening. The reason me going for that approach is after I saw a comment by the competition host.
\n", "rawMarkdown": "I had a model which predicts a very tight bbox around the wheat-heads and it looked very promising but got LB score of lesser than `.6`, it was really saddening. The reason me going for that approach is after I saw a comment by the competition host. \n> In the test set, the bbox are tight to the wheat head (you need to trust me on this one ^^).\nhttps://www.kaggle.com/c/global-wheat-detection/discussion/150781#896963", "votes": null }, { "id": "912454", "postDate": "07/02/2020 13:38:46", "content": "In the test set, the bbox are tight to the wheat head (you need to trust me on this one ^^).\n https://www.kaggle.com/c/global-wheat-detection/discussion/150781#896963
\n
How would you train a model to predict very tight boxes? As a lot of the training data has boxes which have a lot of space around each wheat head
", "rawMarkdown": "How would you train a model to predict very tight boxes? As a lot of the training data has boxes which have a lot of space around each wheat head", "votes": null }, { "id": "912763", "postDate": "07/02/2020 17:55:00", "content": "@matthewmasters I trained a small segmentation model (64*64) that crops out only wheat heads given a bounding box. Fine-tuning with less amount of data was enough to get a good result. The next step was iterate over \ntrain.csv -> crop BBOX image -> run_segmentation -> reassign xmin,ymin,xmax,ymax -> repeat\nUsing this correction I trained my detection model. Although the new bbox correction was not perfect but it looked better than the existing one and thus was really positive to climb the LB but unfortunately it didn't. I still have the corrected bbox CSV file, don't know if it is even allowed or worth the share.
Here are some few examples. I do agree that sometimes that bbox closes more to the wheat box leading to miss some area. (BLUE: Original BBox, RED: Corrected BBox)\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1231059%2F1cf99822e075d08d97e997ae857337fb%2Fexample1.png?generation=1593713003593697&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1231059%2F64563d42958799919081a4ff1e639e9f%2Fexample2.png?generation=1593713135880495&alt=media\")
These come from the original images being split up into 1024 patches. I would hope that they at least fix these for the private test set, as they indicated in their last update. Though with the quality of the data being what it is... I wouldn't be too confident that it'll be get fixed properly.
", "votes": null, "replies": [] }, { "id": 912061, "author_name": "bendang", "author_url": "", "post_date": "07/02/2020 07:23:42", "content": "I had a model which predicts a very tight bbox around the wheat-heads and it looked very promising but got LB score of lesser than .6, it was really saddening. The reason me going for that approach is after I saw a comment by the competition host.
\n", "votes": null, "replies": [ { "id": 912454, "author_name": "matthewmasters", "author_url": "", "post_date": "07/02/2020 13:38:46", "content": "In the test set, the bbox are tight to the wheat head (you need to trust me on this one ^^).\n https://www.kaggle.com/c/global-wheat-detection/discussion/150781#896963
\n
How would you train a model to predict very tight boxes? As a lot of the training data has boxes which have a lot of space around each wheat head
", "votes": null, "replies": [] }, { "id": 912763, "author_name": "bendang", "author_url": "", "post_date": "07/02/2020 17:55:00", "content": "@matthewmasters I trained a small segmentation model (64*64) that crops out only wheat heads given a bounding box. Fine-tuning with less amount of data was enough to get a good result. The next step was iterate over \ntrain.csv -> crop BBOX image -> run_segmentation -> reassign xmin,ymin,xmax,ymax -> repeat\nUsing this correction I trained my detection model. Although the new bbox correction was not perfect but it looked better than the existing one and thus was really positive to climb the LB but unfortunately it didn't. I still have the corrected bbox CSV file, don't know if it is even allowed or worth the share.
Here are some few examples. I do agree that sometimes that bbox closes more to the wheat box leading to miss some area. (BLUE: Original BBox, RED: Corrected BBox)\n\n