![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1856617%2F9201d06503387e01b0777afad098cc9e%2FImage%2004-01-26%20at%201.49PM.jpeg?generation=1767514879011038&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1856617%2F12b7f8f9dadaf6bc49ef5ea722091a0f%2FImage%2004-01-26%20at%201.50PM.jpeg?generation=1767514893944921&alt=media\")
Find the difference between these images one 10gms of dead and other is 50 gms , forget the models even at human level without labels it is hard distinguish between these kind of images!!!
\n
Find the difference between these images one 10gms of dead and other is 50 gms , forget the models even at human level without labels it is hard distinguish between these kind of images!!!
\n
Just to make sure: the first one has more dead matter, right? π
", "rawMarkdown": "Just to make sure: the first one has more dead matter, right? π", "votes": 1 }, { "id": 3386028, "postDate": "2026-01-04T12:23:59.677Z", "content": "The \"dead matter\" is probably hidden underneath the grass which we can't see from the image π€£
", "rawMarkdown": "The \"dead matter\" is probably hidden underneath the grass which we can't see from the image π€£", "votes": 2 }, { "id": 3386316, "postDate": "2026-01-05T03:59:41.993Z", "content": "What about you? How do you process it? @vjagannath786
", "rawMarkdown": "What about you? How do you process it? @vjagannath786 " }, { "id": 3385914, "postDate": "2026-01-04T08:30:01.277Z", "content": "There are indeed so many of such instances which clearly makes this competition all about data enginnering rather than machine learning. Infact, I now regret doing so many model related experiments when I should have been only playing with the data.
", "rawMarkdown": "There are indeed so many of such instances which clearly makes this competition all about data enginnering rather than machine learning. Infact, I now regret doing so many model related experiments when I should have been only playing with the data. ", "replies": [ { "id": 3385917, "postDate": "2026-01-04T08:42:52.107Z", "content": "even if we do data engineering metadata should be available in test set, otherwise there lot of noise. \nJust like this model gets confused whether it is ryegrass or Fescue or the height itself.
\nBtw, did you get any help adding the metadata. I tried adding it as aux loss but it didn't really help in LB.
", "rawMarkdown": "even if we do data engineering metadata should be available in test set, otherwise there lot of noise. \nJust like this model gets confused whether it is ryegrass or Fescue or the height itself. \n\nBtw, did you get any help adding the metadata. I tried adding it as aux loss but it didn't really help in LB.", "replies": [ { "id": 3385923, "postDate": "2026-01-04T09:00:41.170Z", "content": "No, my current best model is not using any meta data.
", "rawMarkdown": "No, my current best model is not using any meta data. ", "votes": 2 } ] }, { "id": 3386030, "postDate": "2026-01-04T12:28:25.873Z", "content": "same π€£ - I have always noticed that model-based solutions to a competition regardless of data gets more attention then data cleaning or data engineering based approaches. It feels more \"impressive\" even if its counter-intuitive π
", "rawMarkdown": "same π€£ - I have always noticed that model-based solutions to a competition regardless of data gets more attention then data cleaning or data engineering based approaches. It feels more \"impressive\" even if its counter-intuitive π", "votes": 1 }, { "id": 3386032, "postDate": "2026-01-04T12:36:37.517Z", "content": "I have also been trying to pivot to more data engineering approaches for the same reason. Unfortunately, this is not a strong suit of mine π
", "rawMarkdown": "I have also been trying to pivot to more data engineering approaches for the same reason. Unfortunately, this is not a strong suit of mine π ", "replies": [ { "id": 3386036, "postDate": "2026-01-04T12:52:18.580Z", "content": "opposite for me, I wanted to gain experience in computer vision tasks haha. Lots of my previous comps are heavy data engineering and I felt I haven't done enough modeling π€£
", "rawMarkdown": "opposite for me, I wanted to gain experience in computer vision tasks haha. Lots of my previous comps are heavy data engineering and I felt I haven't done enough modeling π€£" }, { "id": 3386090, "postDate": "2026-01-04T15:36:44.080Z", "content": "At least the compute requirement was low so it was easy to try many models/approaches!
", "rawMarkdown": "At least the compute requirement was low so it was easy to try many models/approaches!" } ] }, { "id": 3386253, "postDate": "2026-01-04T22:58:41.113Z", "content": "But what kind of data engineering could even out things like this? If visually identical images have wildly different labels?
", "rawMarkdown": "But what kind of data engineering could even out things like this? If visually identical images have wildly different labels?" }, { "id": 3386641, "postDate": "2026-01-05T15:42:04.417Z", "content": "Does data-oriented work include data augmentation? I've been improving the model structure recently, but it's hard to change the LB.πIt also seems that the results obtained through different training methods vary greatly.
", "rawMarkdown": "Does data-oriented work include data augmentation? I've been improving the model structure recently, but it's hard to change the LB.πIt also seems that the results obtained through different training methods vary greatly." }, { "id": 3386642, "postDate": "2026-01-05T15:43:04.583Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 3389999, "author_name": "clemologe", "author_url": "", "post_date": "2026-01-12T11:25:30.787000", "content": "Just to make sure: the first one has more dead matter, right? π
", "votes": 1, "replies": [] }, { "id": 3386028, "author_name": "Geremie Yeo", "author_url": "", "post_date": "2026-01-04T12:23:59.677000", "content": "The \"dead matter\" is probably hidden underneath the grass which we can't see from the image π€£
", "votes": 2, "replies": [] }, { "id": 3386316, "author_name": "Navneet", "author_url": "", "post_date": "2026-01-05T03:59:41.993000", "content": "What about you? How do you process it? @vjagannath786
", "votes": 0, "replies": [] }, { "id": 3385914, "author_name": "Komil Parmar", "author_url": "", "post_date": "2026-01-04T08:30:01.277000", "content": "There are indeed so many of such instances which clearly makes this competition all about data enginnering rather than machine learning. Infact, I now regret doing so many model related experiments when I should have been only playing with the data.
", "votes": 0, "replies": [ { "id": 3385917, "author_name": "Varun Jagannath", "author_url": "", "post_date": "2026-01-04T08:42:52.107000", "content": "even if we do data engineering metadata should be available in test set, otherwise there lot of noise. \nJust like this model gets confused whether it is ryegrass or Fescue or the height itself.
\nBtw, did you get any help adding the metadata. I tried adding it as aux loss but it didn't really help in LB.
", "votes": 0, "replies": [ { "id": 3385923, "author_name": "Komil Parmar", "author_url": "", "post_date": "2026-01-04T09:00:41.170000", "content": "No, my current best model is not using any meta data.
", "votes": 2, "replies": [] } ] }, { "id": 3386030, "author_name": "Geremie Yeo", "author_url": "", "post_date": "2026-01-04T12:28:25.873000", "content": "same π€£ - I have always noticed that model-based solutions to a competition regardless of data gets more attention then data cleaning or data engineering based approaches. It feels more \"impressive\" even if its counter-intuitive π
", "votes": 1, "replies": [] }, { "id": 3386032, "author_name": "chemdatafarmer", "author_url": "", "post_date": "2026-01-04T12:36:37.517000", "content": "I have also been trying to pivot to more data engineering approaches for the same reason. Unfortunately, this is not a strong suit of mine π
", "votes": 0, "replies": [ { "id": 3386036, "author_name": "Geremie Yeo", "author_url": "", "post_date": "2026-01-04T12:52:18.580000", "content": "opposite for me, I wanted to gain experience in computer vision tasks haha. Lots of my previous comps are heavy data engineering and I felt I haven't done enough modeling π€£
", "votes": 0, "replies": [] }, { "id": 3386090, "author_name": "chemdatafarmer", "author_url": "", "post_date": "2026-01-04T15:36:44.080000", "content": "At least the compute requirement was low so it was easy to try many models/approaches!
", "votes": 0, "replies": [] } ] }, { "id": 3386253, "author_name": "thomas rost", "author_url": "", "post_date": "2026-01-04T22:58:41.113000", "content": "But what kind of data engineering could even out things like this? If visually identical images have wildly different labels?
", "votes": 0, "replies": [] }, { "id": 3386641, "author_name": "YYH", "author_url": "", "post_date": "2026-01-05T15:42:04.417000", "content": "Does data-oriented work include data augmentation? I've been improving the model structure recently, but it's hard to change the LB.πIt also seems that the results obtained through different training methods vary greatly.
", "votes": 0, "replies": [] }, { "id": 3386642, "author_name": "", "author_url": "", "post_date": "2026-01-05T15:43:04.583000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "3385913": "Find the difference between these images one 10gms of dead and other is 50 gms , forget the models even at human level without labels it is hard distinguish between these kind of images!!!\n\n", "3389999": "Just to make sure: the first one has more dead matter, right? π", "3386028": "The \"dead matter\" is probably hidden underneath the grass which we can't see from the image π€£", "3386316": "What about you? How do you process it? @vjagannath786 ", "3385914": "There are indeed so many of such instances which clearly makes this competition all about data enginnering rather than machine learning. Infact, I now regret doing so many model related experiments when I should have been only playing with the data. " } }