The training csv has the metadata but the test csv does not? So what is the point of using the metadata for training?
", "messages": [ { "id": "3332453", "postDate": "11/16/2025 17:49:30", "content": "The training csv has the metadata but the test csv does not? So what is the point of using the metadata for training?
", "rawMarkdown": "The training csv has the metadata but the test csv does not? So what is the point of using the metadata for training?", "votes": null }, { "id": "3332490", "postDate": "11/16/2025 18:06:18", "content": "If you read the paper that describes the data collection a decade ago you might get the impression (as I did) that the experiment was focused on using the metadata only for generating the models. The images seemed like mostly only intended for quality control - which I think explains why so many have edges of note cards, shadows of the camera stand, people shadows, etc. I believe they found at least half of the images that had been taken to be not suitable for this competition.
\nSo pretty sure that the metadata is available (not to us:) for both the train and the test images.
\nThey apparently developed an app and I assume used it. I assume that the app required users to collect metadata.
\nIf you read the paper, the data collection for the meta data would likely be a real pain in the arse for a rancher to do with any precision or speed.
\nI spent a decent amount of time a few years ago on a 25,000 acre cattle ranch were the rancher needed to pay close attention to the grass status. It was so critical in fact that he spent lots of dollars getting insurance that covered his losses if the grass was really poor in a given year. There was a lot of variation within each designated grazing lot, so a decent number of sample collections would be required for a good understanding of the status of that lot.
\nThe expectation that a rancher would have the resources to collect and dry biomass to use an app would be almost certainly a reason for it to fail to catch on as a handy useful tool.
\nSo real life use - you would want to use only images, and collect a bunch of them, probably with a camera mounted on an off road quad vehicle.
\nThe meta existed - so why not provide it. I have built a model that predicts the meta from images and could use that to predict the test images and therefore have use of the meta for train and test. There is a discussion post somewhere where someone has suggested that the top scoring models are using images and meta data predictions.
\nJust my opinion - but think I have high probability of getting most of the comments right.
", "rawMarkdown": "If you read the paper that describes the data collection a decade ago you might get the impression (as I did) that the experiment was focused on using the metadata only for generating the models. The images seemed like mostly only intended for quality control - which I think explains why so many have edges of note cards, shadows of the camera stand, people shadows, etc. I believe they found at least half of the images that had been taken to be not suitable for this competition.\n\nSo pretty sure that the metadata is available (not to us:) for both the train and the test images.\n\nThey apparently developed an app and I assume used it. I assume that the app required users to collect metadata.\n\nIf you read the paper, the data collection for the meta data would likely be a real pain in the arse for a rancher to do with any precision or speed. \n\nI spent a decent amount of time a few years ago on a 25,000 acre cattle ranch were the rancher needed to pay close attention to the grass status. It was so critical in fact that he spent lots of dollars getting insurance that covered his losses if the grass was really poor in a given year. There was a lot of variation within each designated grazing lot, so a decent number of sample collections would be required for a good understanding of the status of that lot. \n\nThe expectation that a rancher would have the resources to collect and dry biomass to use an app would be almost certainly a reason for it to fail to catch on as a handy useful tool.\n\nSo real life use - you would want to use only images, and collect a bunch of them, probably with a camera mounted on an off road quad vehicle.\n\nThe meta existed - so why not provide it. I have built a model that predicts the meta from images and could use that to predict the test images and therefore have use of the meta for train and test. There is a discussion post somewhere where someone has suggested that the top scoring models are using images and meta data predictions.\n\nJust my opinion - but think I have high probability of getting most of the comments right.", "votes": null }, { "id": "3332812", "postDate": "11/17/2025 00:43:34", "content": "Thx Jimmy, Always learning very much from you😀
", "rawMarkdown": "Thx Jimmy, Always learning very much from you😀", "votes": null } ], "comments": [ { "id": 3332490, "author_name": "pcjimmmy", "author_url": "", "post_date": "11/16/2025 18:06:18", "content": "If you read the paper that describes the data collection a decade ago you might get the impression (as I did) that the experiment was focused on using the metadata only for generating the models. The images seemed like mostly only intended for quality control - which I think explains why so many have edges of note cards, shadows of the camera stand, people shadows, etc. I believe they found at least half of the images that had been taken to be not suitable for this competition.
\nSo pretty sure that the metadata is available (not to us:) for both the train and the test images.
\nThey apparently developed an app and I assume used it. I assume that the app required users to collect metadata.
\nIf you read the paper, the data collection for the meta data would likely be a real pain in the arse for a rancher to do with any precision or speed.
\nI spent a decent amount of time a few years ago on a 25,000 acre cattle ranch were the rancher needed to pay close attention to the grass status. It was so critical in fact that he spent lots of dollars getting insurance that covered his losses if the grass was really poor in a given year. There was a lot of variation within each designated grazing lot, so a decent number of sample collections would be required for a good understanding of the status of that lot.
\nThe expectation that a rancher would have the resources to collect and dry biomass to use an app would be almost certainly a reason for it to fail to catch on as a handy useful tool.
\nSo real life use - you would want to use only images, and collect a bunch of them, probably with a camera mounted on an off road quad vehicle.
\nThe meta existed - so why not provide it. I have built a model that predicts the meta from images and could use that to predict the test images and therefore have use of the meta for train and test. There is a discussion post somewhere where someone has suggested that the top scoring models are using images and meta data predictions.
\nJust my opinion - but think I have high probability of getting most of the comments right.
", "votes": null, "replies": [ { "id": 3332812, "author_name": "qiyuliaocsiro", "author_url": "", "post_date": "11/17/2025 00:43:34", "content": "Thx Jimmy, Always learning very much from you😀
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3332453": "The training csv has the metadata but the test csv does not? So what is the point of using the metadata for training?", "3332490": "If you read the paper that describes the data collection a decade ago you might get the impression (as I did) that the experiment was focused on using the metadata only for generating the models. The images seemed like mostly only intended for quality control - which I think explains why so many have edges of note cards, shadows of the camera stand, people shadows, etc. I believe they found at least half of the images that had been taken to be not suitable for this competition.\n\nSo pretty sure that the metadata is available (not to us:) for both the train and the test images.\n\nThey apparently developed an app and I assume used it. I assume that the app required users to collect metadata.\n\nIf you read the paper, the data collection for the meta data would likely be a real pain in the arse for a rancher to do with any precision or speed. \n\nI spent a decent amount of time a few years ago on a 25,000 acre cattle ranch were the rancher needed to pay close attention to the grass status. It was so critical in fact that he spent lots of dollars getting insurance that covered his losses if the grass was really poor in a given year. There was a lot of variation within each designated grazing lot, so a decent number of sample collections would be required for a good understanding of the status of that lot. \n\nThe expectation that a rancher would have the resources to collect and dry biomass to use an app would be almost certainly a reason for it to fail to catch on as a handy useful tool.\n\nSo real life use - you would want to use only images, and collect a bunch of them, probably with a camera mounted on an off road quad vehicle.\n\nThe meta existed - so why not provide it. I have built a model that predicts the meta from images and could use that to predict the test images and therefore have use of the meta for train and test. There is a discussion post somewhere where someone has suggested that the top scoring models are using images and meta data predictions.\n\nJust my opinion - but think I have high probability of getting most of the comments right.", "3332812": "Thx Jimmy, Always learning very much from you😀" }, "source": "meta" }