{ "id": 309121, "title": "Another Image Detection & Crop", "url": "/competitions/happy-whale-and-dolphin/discussion/309121", "author_name": "", "post_date": "2022-02-22T00:33:05.841301Z", "votes": 14, "comment_count": 7, "views": 0, "content": "

Hi peoples, I'm new here. I hand bboxed some images ~800 of the training and ran it through YOLOv5. I intend to crop and resize, I see others methods have used 512x512. Is that a good size?

\n

It may not do better than the existing ones out there, but it's fun to try.

\n

It's not perfect by the way, maybe i should tag more and try again, it really didn't take long to tag, it took like one terrible netflix movie.

\n


\n

", "messages": [ { "id": "1700414", "postDate": "02/22/2022 00:33:05", "content": "

Hi peoples, I'm new here. I hand bboxed some images ~800 of the training and ran it through YOLOv5. I intend to crop and resize, I see others methods have used 512x512. Is that a good size?

\n

It may not do better than the existing ones out there, but it's fun to try.

\n

It's not perfect by the way, maybe i should tag more and try again, it really didn't take long to tag, it took like one terrible netflix movie.

\n


\n

", "rawMarkdown": "Hi peoples, I'm new here. I hand bboxed some images ~800 of the training and ran it through YOLOv5. I intend to crop and resize, I see others methods have used 512x512. Is that a good size?\n\nIt may not do better than the existing ones out there, but it's fun to try.\n\nIt's not perfect by the way, maybe i should tag more and try again, it really didn't take long to tag, it took like one terrible netflix movie.\n\n\n", "votes": null }, { "id": "1700680", "postDate": "02/22/2022 06:38:23", "content": "

Nice job! I believe we can combine it with detic and yolo. Do you plan to publish your work?

", "rawMarkdown": "Nice job! I believe we can combine it with detic and yolo. Do you plan to publish your work?", "votes": null }, { "id": "1701434", "postDate": "02/22/2022 19:17:55", "content": "

Do you mean share the pt file or the yolo txt files? I can do that, I labeled more yesterday, about ~1400 total, and it got better, but I think it can do even better, so I'm going to label some more today, and then I will post them and the pt file if people don't want to train.
\nI would think optimally we want the most visible amount of the \"primary\" animal above water.
\nSo I don't know what is the best cropping algorithm to do this, e.g. based on detection box and probability.

\n

Anyway, I am new to this so my labeling skills are probably subpar, I did not strategically pic photos to label, and I honestly don't know what the best parameters are for training yolov5 other than copying mostly from @awsaf49

", "rawMarkdown": "Do you mean share the pt file or the yolo txt files? I can do that, I labeled more yesterday, about ~1400 total, and it got better, but I think it can do even better, so I'm going to label some more today, and then I will post them and the pt file if people don't want to train.\nI would think optimally we want the most visible amount of the \"primary\" animal above water.\nSo I don't know what is the best cropping algorithm to do this, e.g. based on detection box and probability.\n\nAnyway, I am new to this so my labeling skills are probably subpar, I did not strategically pic photos to label, and I honestly don't know what the best parameters are for training yolov5 other than copying mostly from @awsaf49", "votes": null }, { "id": "1701601", "postDate": "02/22/2022 23:28:13", "content": "

Great effort! 🙏

", "rawMarkdown": "Great effort! 🙏", "votes": null }, { "id": "1701831", "postDate": "02/23/2022 05:07:49", "content": "

Thanks for sharing, here are another 1500 annotations, sampling 3% from all species https://www.kaggle.com/bsridatta/happierwhale-1500-manual-annotations

", "rawMarkdown": "Thanks for sharing, here are another 1500 annotations, sampling 3% from all species https://www.kaggle.com/bsridatta/happierwhale-1500-manual-annotations", "votes": null }, { "id": "1702712", "postDate": "02/23/2022 20:53:17", "content": "

Thanks for sharing!
\nMaybe you could fine tune the other previously shared detic or yolo models with your data and see if you can improve them?

", "rawMarkdown": "Thanks for sharing! \nMaybe you could fine tune the other previously shared detic or yolo models with your data and see if you can improve them?", "votes": null }, { "id": "1702790", "postDate": "02/23/2022 23:05:04", "content": "

son of a narwhal… awesome thank you, probably lot's of overlap, I have 2200 labeled

", "rawMarkdown": "son of a narwhal... awesome thank you, probably lot's of overlap, I have 2200 labeled", "votes": null }, { "id": "1702802", "postDate": "02/23/2022 23:50:27", "content": "

NM, Not a lot of overlap, 3,648 unique

", "rawMarkdown": "NM, Not a lot of overlap, 3,648 unique", "votes": null } ], "comments": [ { "id": 1700680, "author_name": "meowmeowmeowmeowmeow", "author_url": "", "post_date": "02/22/2022 06:38:23", "content": "

Nice job! I believe we can combine it with detic and yolo. Do you plan to publish your work?

", "votes": null, "replies": [ { "id": 1701434, "author_name": "click2clone", "author_url": "", "post_date": "02/22/2022 19:17:55", "content": "

Do you mean share the pt file or the yolo txt files? I can do that, I labeled more yesterday, about ~1400 total, and it got better, but I think it can do even better, so I'm going to label some more today, and then I will post them and the pt file if people don't want to train.
\nI would think optimally we want the most visible amount of the \"primary\" animal above water.
\nSo I don't know what is the best cropping algorithm to do this, e.g. based on detection box and probability.

\n

Anyway, I am new to this so my labeling skills are probably subpar, I did not strategically pic photos to label, and I honestly don't know what the best parameters are for training yolov5 other than copying mostly from @awsaf49

", "votes": null, "replies": [] } ] }, { "id": 1701601, "author_name": "rajborgohain", "author_url": "", "post_date": "02/22/2022 23:28:13", "content": "

Great effort! 🙏

", "votes": null, "replies": [] }, { "id": 1701831, "author_name": "bsridatta", "author_url": "", "post_date": "02/23/2022 05:07:49", "content": "

Thanks for sharing, here are another 1500 annotations, sampling 3% from all species https://www.kaggle.com/bsridatta/happierwhale-1500-manual-annotations

", "votes": null, "replies": [ { "id": 1702790, "author_name": "click2clone", "author_url": "", "post_date": "02/23/2022 23:05:04", "content": "

son of a narwhal… awesome thank you, probably lot's of overlap, I have 2200 labeled

", "votes": null, "replies": [] }, { "id": 1702802, "author_name": "click2clone", "author_url": "", "post_date": "02/23/2022 23:50:27", "content": "

NM, Not a lot of overlap, 3,648 unique

", "votes": null, "replies": [] } ] }, { "id": 1702712, "author_name": "ashwincheekati", "author_url": "", "post_date": "02/23/2022 20:53:17", "content": "

Thanks for sharing!
\nMaybe you could fine tune the other previously shared detic or yolo models with your data and see if you can improve them?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1700414": "Hi peoples, I'm new here. I hand bboxed some images ~800 of the training and ran it through YOLOv5. I intend to crop and resize, I see others methods have used 512x512. Is that a good size?\n\nIt may not do better than the existing ones out there, but it's fun to try.\n\nIt's not perfect by the way, maybe i should tag more and try again, it really didn't take long to tag, it took like one terrible netflix movie.\n\n\n", "1700680": "Nice job! I believe we can combine it with detic and yolo. Do you plan to publish your work?", "1701434": "Do you mean share the pt file or the yolo txt files? I can do that, I labeled more yesterday, about ~1400 total, and it got better, but I think it can do even better, so I'm going to label some more today, and then I will post them and the pt file if people don't want to train.\nI would think optimally we want the most visible amount of the \"primary\" animal above water.\nSo I don't know what is the best cropping algorithm to do this, e.g. based on detection box and probability.\n\nAnyway, I am new to this so my labeling skills are probably subpar, I did not strategically pic photos to label, and I honestly don't know what the best parameters are for training yolov5 other than copying mostly from @awsaf49", "1701601": "Great effort! 🙏", "1701831": "Thanks for sharing, here are another 1500 annotations, sampling 3% from all species https://www.kaggle.com/bsridatta/happierwhale-1500-manual-annotations", "1702712": "Thanks for sharing! \nMaybe you could fine tune the other previously shared detic or yolo models with your data and see if you can improve them?", "1702790": "son of a narwhal... awesome thank you, probably lot's of overlap, I have 2200 labeled", "1702802": "NM, Not a lot of overlap, 3,648 unique" }, "source": "meta" }