![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1994602%2F417e33a82e6b3ea4df8a63adebc7b3b7%2F__results___9_3.png?generation=1593587072577241&alt=media\")
Using worldwide data, you will focus on a generalized solution to estimate the number and size of wheat heads. To better gauge the performance for unseen genotypes, environments, and observational conditions, the training dataset covers multiple regions. You will use more than 3,000 images from Europe (France, UK, Switzerland) and North America (Canada). The test data includes about 1,000 images from Australia, Japan, and China.
What this means is that the model is going to see samples that may be totally new to it.
\n\nTo simulate this, I tried to split the data according to location, and select one of the folds which seem to have good train-valid distribution. It can serve as a good testbench to try out different generalization strategies.
\n\nHere's the generated dataset, and here's the notebook I used to generate the data.
\n\nI am thinking of using RRES as the validation set and the others as the training set. What do you guys think?\n
Using worldwide data, you will focus on a generalized solution to estimate the number and size of wheat heads. To better gauge the performance for unseen genotypes, environments, and observational conditions, the training dataset covers multiple regions. You will use more than 3,000 images from Europe (France, UK, Switzerland) and North America (Canada). The test data includes about 1,000 images from Australia, Japan, and China.
What this means is that the model is going to see samples that may be totally new to it.
\n\nTo simulate this, I tried to split the data according to location, and select one of the folds which seem to have good train-valid distribution. It can serve as a good testbench to try out different generalization strategies.
\n\nHere's the generated dataset, and here's the notebook I used to generate the data.
\n\nI am thinking of using RRES as the validation set and the others as the training set. What do you guys think?\n
Interesting!
", "rawMarkdown": "Interesting!", "votes": null }, { "id": "911188", "postDate": "07/01/2020 15:51:50", "content": "I also thought this would be a good approach at the start of the competition. Though I have kindof changed my thoughts on it and believe including as diverse a validation set as possible is actually the best approach. See more discussion on this here.
", "rawMarkdown": "I also thought this would be a good approach at the start of the competition. Though I have kindof changed my thoughts on it and believe including as diverse a validation set as possible is actually the best approach. See more discussion on this [here](https://www.kaggle.com/c/global-wheat-detection/discussion/152569#855752).", "votes": null }, { "id": "911205", "postDate": "07/01/2020 16:00:50", "content": "Makes sense. Though what I had in my mind was to train a cycleGAN to see if it can generate new samples to simulate the unseen data.
", "rawMarkdown": "Makes sense. Though what I had in my mind was to train a cycleGAN to see if it can generate new samples to simulate the unseen data.", "votes": null }, { "id": "911221", "postDate": "07/01/2020 16:03:58", "content": "Yeah... I've actually tried something like that too to get over the domain generalization problem. Unfortunately I didn't have much luck and wasted a bit of time but I wish you the best of luck! :D
", "rawMarkdown": "Yeah... I've actually tried something like that too to get over the domain generalization problem. Unfortunately I didn't have much luck and wasted a bit of time but I wish you the best of luck! :D", "votes": null }, { "id": "911238", "postDate": "07/01/2020 16:08:30", "content": "Thanks! I've been looking for a push to try out cycleGAN for a long time. I'll be sure to update you if I get any positive results :D\nAll the best to you too!
", "rawMarkdown": "Thanks! I've been looking for a push to try out cycleGAN for a long time. I'll be sure to update you if I get any positive results :D\nAll the best to you too!", "votes": null } ], "comments": [ { "id": 910388, "author_name": "jonykarki", "author_url": "", "post_date": "07/01/2020 06:06:36", "content": "Interesting!
", "votes": null, "replies": [] }, { "id": 911188, "author_name": "matthewmasters", "author_url": "", "post_date": "07/01/2020 15:51:50", "content": "I also thought this would be a good approach at the start of the competition. Though I have kindof changed my thoughts on it and believe including as diverse a validation set as possible is actually the best approach. See more discussion on this here.
", "votes": null, "replies": [ { "id": 911205, "author_name": "mightyrains", "author_url": "", "post_date": "07/01/2020 16:00:50", "content": "Makes sense. Though what I had in my mind was to train a cycleGAN to see if it can generate new samples to simulate the unseen data.
", "votes": null, "replies": [] }, { "id": 911221, "author_name": "matthewmasters", "author_url": "", "post_date": "07/01/2020 16:03:58", "content": "Yeah... I've actually tried something like that too to get over the domain generalization problem. Unfortunately I didn't have much luck and wasted a bit of time but I wish you the best of luck! :D
", "votes": null, "replies": [] }, { "id": 911238, "author_name": "mightyrains", "author_url": "", "post_date": "07/01/2020 16:08:30", "content": "Thanks! I've been looking for a push to try out cycleGAN for a long time. I'll be sure to update you if I get any positive results :D\nAll the best to you too!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "910214": "`Using worldwide data, you will focus on a generalized solution to estimate the number and size of wheat heads. To better gauge the performance for unseen genotypes, environments, and observational conditions, the training dataset covers multiple regions. You will use more than 3,000 images from Europe (France, UK, Switzerland) and North America (Canada). The test data includes about 1,000 images from Australia, Japan, and China.`\n\nWhat this means is that the model is going to see samples that may be totally new to it.\n\nTo simulate this, I tried to split the data according to location, and select one of the folds which seem to have good train-valid distribution. It can serve as a good testbench to try out different generalization strategies. \n\n[Here's](https://www.kaggle.com/mightyrains/wheat-split-by-location) the generated dataset, and [here's](https://www.kaggle.com/mightyrains/wheat-eda-location-wise-split) the notebook I used to generate the data.\n\nI am thinking of using RRES as the validation set and the others as the training set. What do you guys think?\n", "910388": "Interesting!", "911188": "I also thought this would be a good approach at the start of the competition. Though I have kindof changed my thoughts on it and believe including as diverse a validation set as possible is actually the best approach. See more discussion on this [here](https://www.kaggle.com/c/global-wheat-detection/discussion/152569#855752).", "911205": "Makes sense. Though what I had in my mind was to train a cycleGAN to see if it can generate new samples to simulate the unseen data.", "911221": "Yeah... I've actually tried something like that too to get over the domain generalization problem. Unfortunately I didn't have much luck and wasted a bit of time but I wish you the best of luck! :D", "911238": "Thanks! I've been looking for a push to try out cycleGAN for a long time. I'll be sure to update you if I get any positive results :D\nAll the best to you too!" }, "source": "meta" }