I'm maintaining a GitHub repository having all the code I've written for this competition with approaches and their results documented. Here it is: https://github.com/Kaushal28/Global-Wheat-Detection
\n\nHope it's helpful to someone! Thanks!
", "messages": [ { "id": "847989", "postDate": "05/14/2020 17:34:11", "content": "I'm maintaining a GitHub repository having all the code I've written for this competition with approaches and their results documented. Here it is: https://github.com/Kaushal28/Global-Wheat-Detection
\n\nHope it's helpful to someone! Thanks!
", "rawMarkdown": "I'm maintaining a GitHub repository having all the code I've written for this competition with approaches and their results documented. Here it is: https://github.com/Kaushal28/Global-Wheat-Detection\n\nHope it's helpful to someone! Thanks!", "votes": null }, { "id": "848114", "postDate": "05/14/2020 18:40:56", "content": "Thanks for sharing :D
", "rawMarkdown": "Thanks for sharing :D", "votes": null }, { "id": "848995", "postDate": "05/15/2020 12:11:41", "content": "@kaushal2896 very useful as always. Thank you!\nI have no idea why but I'm not able to get as high as everyone else with faster rcnn. Perhaps it's because I'm using a different split for train-val.\nSo two questions to make sure I am doing the same thing as you:
\n\n@alexandersoare I'm also not able to get even 0.66 with faster RCNN and pseudo labeling. The one thing I have not tried is data augmentations and I think that can be the main reason. Anyways answer to your questions:
\n\nYes I'm using faster RCNN with FPN. You can check PyTorch docs to confirm it. I've referred this to write my training script: https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html
My results are based off training on all the data without pseudo labels and training on pseudo labels for one, two and three rounds. I've not tried ensemble yet.
I think the best approach is to develop your models using a good train/valid split however, once you've improved your model and want to submit it, use all data in order to train.
", "rawMarkdown": "I think the best approach is to develop your models using a good train/valid split however, once you've improved your model and want to submit it, use all data in order to train.", "votes": null }, { "id": "849898", "postDate": "05/16/2020 06:53:25", "content": "Only fasterRcnn could get 162nd ? \nThen CascadRcnn should be better.
", "rawMarkdown": "Only fasterRcnn could get 162nd ? \nThen CascadRcnn should be better.", "votes": null } ], "comments": [ { "id": 848114, "author_name": "albeffe", "author_url": "", "post_date": "05/14/2020 18:40:56", "content": "Thanks for sharing :D
", "votes": null, "replies": [] }, { "id": 848995, "author_name": "alexandersoare", "author_url": "", "post_date": "05/15/2020 12:11:41", "content": "@kaushal2896 very useful as always. Thank you!\nI have no idea why but I'm not able to get as high as everyone else with faster rcnn. Perhaps it's because I'm using a different split for train-val.\nSo two questions to make sure I am doing the same thing as you:
\n\n@alexandersoare I'm also not able to get even 0.66 with faster RCNN and pseudo labeling. The one thing I have not tried is data augmentations and I think that can be the main reason. Anyways answer to your questions:
\n\nYes I'm using faster RCNN with FPN. You can check PyTorch docs to confirm it. I've referred this to write my training script: https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html
My results are based off training on all the data without pseudo labels and training on pseudo labels for one, two and three rounds. I've not tried ensemble yet.
I think the best approach is to develop your models using a good train/valid split however, once you've improved your model and want to submit it, use all data in order to train.
", "votes": null, "replies": [] } ] }, { "id": 849898, "author_name": "xujingzhao", "author_url": "", "post_date": "05/16/2020 06:53:25", "content": "Only fasterRcnn could get 162nd ? \nThen CascadRcnn should be better.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "847989": "I'm maintaining a GitHub repository having all the code I've written for this competition with approaches and their results documented. Here it is: https://github.com/Kaushal28/Global-Wheat-Detection\n\nHope it's helpful to someone! Thanks!", "848114": "Thanks for sharing :D", "848995": "kaushal2896 very useful as always. Thank you!\nI have no idea why but I'm not able to get as high as everyone else with faster rcnn. Perhaps it's because I'm using a different split for train-val.\nSo two questions to make sure I am doing the same thing as you:\n\n1. Is this Faster-RCNN with FPN? Or just vanilla?\n2. Are these results based off training on all the data? Training on a split? Or ensembled over multiple splits?", "849821": "alexandersoare I'm also not able to get even 0.66 with faster RCNN and pseudo labeling. The one thing I have not tried is data augmentations and I think that can be the main reason. Anyways answer to your questions:\n\n1. Yes I'm using faster RCNN with FPN. You can check PyTorch docs to confirm it. I've referred this to write my training script: https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html\n\n2. My results are based off training on all the data without pseudo labels and training on pseudo labels for one, two and three rounds. I've not tried ensemble yet.", "849829": "I think the best approach is to develop your models using a good train/valid split however, once you've improved your model and want to submit it, use all data in order to train.", "849898": "Only fasterRcnn could get 162nd ? \nThen CascadRcnn should be better." }, "source": "meta" }