![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4233602%2F221423bbd9ac6165ecc569966a65d217%2F0cbe3f32c.jpeg?generation=1596382851094515&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4233602%2Ff1674b4fa619785a9db28578546ecec2%2F0ce3a22a3.jpeg?generation=1596382888524858&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4233602%2Fef56c68407e83da8c91092c2eae1c43d%2F0cf7ef43d.jpeg?generation=1596382888526925&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4233602%2F9c423ffdc83c63a6aadb8df13740f5f4%2F0d2948e6d.jpeg?generation=1596382888612228&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4233602%2Fb3d3b3212727206a5b27a2f12dc8e229%2F0d2e2b65b.jpeg?generation=1596382888727502&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F4233602%2F077f48e5e5ebea77dd102a35372ac48b%2F0d82ef017.jpeg?generation=1596382890000862&alt=media\")
Hi, I've been comparing my models' predictions to the ground truths and I've been getting a lot of false false positives, meaning that my model found wheat heads, that were there but are not marked by any bounding boxes. \nThese are typically at the edges of the pictures. This is an extremely common issue. The examples below are from a random group of 10 pictures. So out of 10 pictures 6 had unmarked wheat heads and most had multiple. I am getting a lot of false positives for finding wheat heads that are not marked. Especially on the picture that only has one wheat head, which is unmarked, I would have gotten an accuracy of 0 for the entire picture.\nSo please if you haven't already, double-check your private dataset, especially the edges. That being said, I am having a lot of fun in this challenge, I just wanted to draw attention to this issue, because I didn't see mention of it in the Competition update post.
\n\n\n\n\n\n\n
Hi, I've been comparing my models' predictions to the ground truths and I've been getting a lot of false false positives, meaning that my model found wheat heads, that were there but are not marked by any bounding boxes. \nThese are typically at the edges of the pictures. This is an extremely common issue. The examples below are from a random group of 10 pictures. So out of 10 pictures 6 had unmarked wheat heads and most had multiple. I am getting a lot of false positives for finding wheat heads that are not marked. Especially on the picture that only has one wheat head, which is unmarked, I would have gotten an accuracy of 0 for the entire picture.\nSo please if you haven't already, double-check your private dataset, especially the edges. That being said, I am having a lot of fun in this challenge, I just wanted to draw attention to this issue, because I didn't see mention of it in the Competition update post.
\n\n\n\n\n\n\n
I agree that labels may not be perfect. According to the paper, however, some wheats were really intended to not label if the basal parts were less than 30% visible, so this could cover some few cases in the above pictures \n(our models are too good to detect very small wheat parts 😁 )
\n\nhttps://arxiv.org/ftp/arxiv/papers/2005/2005.02162.pdf (page 5-6 Figure 2.e )
", "rawMarkdown": "I agree that labels may not be perfect. According to the paper, however, some wheats were really intended to not label if the basal parts were less than 30% visible, so this could cover some few cases in the above pictures \n(our models are too good to detect very small wheat parts 😁 )\n\nhttps://arxiv.org/ftp/arxiv/papers/2005/2005.02162.pdf (page 5-6 Figure 2.e )", "votes": null } ], "comments": [ { "id": 955983, "author_name": "ratthachat", "author_url": "", "post_date": "08/03/2020 06:09:49", "content": "I agree that labels may not be perfect. According to the paper, however, some wheats were really intended to not label if the basal parts were less than 30% visible, so this could cover some few cases in the above pictures \n(our models are too good to detect very small wheat parts 😁 )
\n\nhttps://arxiv.org/ftp/arxiv/papers/2005/2005.02162.pdf (page 5-6 Figure 2.e )
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "955394": "Hi, I've been comparing my models' predictions to the ground truths and I've been getting a lot of false false positives, meaning that my model found wheat heads, that were there but are not marked by any bounding boxes. \nThese are typically at the edges of the pictures. This is an extremely common issue. The examples below are from a random group of 10 pictures. So out of 10 pictures 6 had unmarked wheat heads and most had multiple. I am getting a lot of false positives for finding wheat heads that are not marked. Especially on the picture that only has one wheat head, which is unmarked, I would have gotten an accuracy of 0 for the entire picture.\nSo please if you haven't already, double-check your private dataset, especially the edges. That being said, I am having a lot of fun in this challenge, I just wanted to draw attention to this issue, because I didn't see mention of it in the Competition update post.\n\n\n\n\n\n\n", "955983": "I agree that labels may not be perfect. According to the paper, however, some wheats were really intended to not label if the basal parts were less than 30% visible, so this could cover some few cases in the above pictures \n(our models are too good to detect very small wheat parts 😁 )\n\nhttps://arxiv.org/ftp/arxiv/papers/2005/2005.02162.pdf (page 5-6 Figure 2.e )" }, "source": "meta" }