haar cascade fins detector
\n\n\ncleaning the image data is crucial for proper performance of classifiers!
\n
In order to do so, please find shared in the link a custom haar-cascade filter specifically designed to detect fins of dolphins and whales in the image,
\nhere is the link in my github: https://github.com/JeanMILPIED/happy_whale_and_dolphin_haar/blob/main/README.md
\nyou can use the filter freely and add it in an image treatment loop,
\nthis may sound very old school, but this is much more easy to compute than a yolovX deeplearning detection considering that almost all images have fins.
\nfeel free to comment hereunder
", "messages": [ { "id": "1695993", "postDate": "02/18/2022 14:27:03", "content": "haar cascade fins detector
\n\n\ncleaning the image data is crucial for proper performance of classifiers!
\n
In order to do so, please find shared in the link a custom haar-cascade filter specifically designed to detect fins of dolphins and whales in the image,
\nhere is the link in my github: https://github.com/JeanMILPIED/happy_whale_and_dolphin_haar/blob/main/README.md
\nyou can use the filter freely and add it in an image treatment loop,
\nthis may sound very old school, but this is much more easy to compute than a yolovX deeplearning detection considering that almost all images have fins.
\nfeel free to comment hereunder
", "rawMarkdown": "**haar cascade fins detector**\n> cleaning the image data is crucial for proper performance of classifiers! \n\nIn order to do so, please find shared in the link a custom haar-cascade filter specifically designed to detect fins of dolphins and whales in the image,\n\nhere is the link in my github: https://github.com/JeanMILPIED/happy_whale_and_dolphin_haar/blob/main/README.md\n\nyou can use the filter freely and add it in an image treatment loop,\n\nthis may sound very old school, but this is much more easy to compute than a yolovX deeplearning detection considering that almost all images have fins.\n\nfeel free to comment hereunder", "votes": null }, { "id": "1696419", "postDate": "02/18/2022 20:20:57", "content": "Oh wow, thank you for sharing, @jeanmilpied! Amazing!
", "rawMarkdown": "Oh wow, thank you for sharing, @jeanmilpied! Amazing!", "votes": null }, { "id": "1697635", "postDate": "02/19/2022 18:21:27", "content": "I pushed my f1 score for whale or dolphin prediction from 73% to 80% using the cropped images with my home made haar cascade fins detector. It is definitely not perfect but worth giving it a try
", "rawMarkdown": "I pushed my f1 score for whale or dolphin prediction from 73% to 80% using the cropped images with my home made haar cascade fins detector. It is definitely not perfect but worth giving it a try", "votes": null }, { "id": "1697685", "postDate": "02/19/2022 19:19:22", "content": "How have you managed to deal with bieluga and grey whale?
", "rawMarkdown": "How have you managed to deal with bieluga and grey whale?", "votes": null }, { "id": "1697797", "postDate": "02/19/2022 21:02:54", "content": "As they have no fins, in case no fin is detected, the whole image is kept. Not exactly the perfect solution, but still a good simplification
", "rawMarkdown": "As they have no fins, in case no fin is detected, the whole image is kept. Not exactly the perfect solution, but still a good simplification", "votes": null }, { "id": "1697800", "postDate": "02/19/2022 21:05:35", "content": "Thank you for answer. Could you make screenshot of your background images? As I remember haar needs GT and background images. To be honest I always had problem with background which lead to poor result and this is why I do not use this kind of custom detector anymore.
", "rawMarkdown": "Thank you for answer. Could you make screenshot of your background images? As I remember haar needs GT and background images. To be honest I always had problem with background which lead to poor result and this is why I do not use this kind of custom detector anymore.", "votes": null }, { "id": "1697810", "postDate": "02/19/2022 21:14:29", "content": "For background images for training, I played silly:
\nA 100 images of lanscapes, a 100 scrapped images of sea and ocean and a 100 images of dogs
\nI used cats as landscapes but I thouth their ears really look like fins sometimes
OK. I see. Should work and it works. Great. I had always with this part problem. Thank you for advice.
", "rawMarkdown": "OK. I see. Should work and it works. Great. I had always with this part problem. Thank you for advice.", "votes": null }, { "id": "1711898", "postDate": "03/04/2022 12:51:05", "content": "Dear Jean
\nThank you for your comments. just a question. As I understood, you create your own dataset that contains images with fines, yes? what is your size of images?
True
\nI made a folder with about 250 images of \"only\" fins based on the compet dataset
Oh wow, thank you for sharing, @jeanmilpied! Amazing!
", "votes": null, "replies": [] }, { "id": 1697635, "author_name": "jeanmilpied", "author_url": "", "post_date": "02/19/2022 18:21:27", "content": "I pushed my f1 score for whale or dolphin prediction from 73% to 80% using the cropped images with my home made haar cascade fins detector. It is definitely not perfect but worth giving it a try
", "votes": null, "replies": [ { "id": 1711898, "author_name": "omiderfanmanesh", "author_url": "", "post_date": "03/04/2022 12:51:05", "content": "Dear Jean
\nThank you for your comments. just a question. As I understood, you create your own dataset that contains images with fines, yes? what is your size of images?
True
\nI made a folder with about 250 images of \"only\" fins based on the compet dataset
How have you managed to deal with bieluga and grey whale?
", "votes": null, "replies": [ { "id": 1697797, "author_name": "jeanmilpied", "author_url": "", "post_date": "02/19/2022 21:02:54", "content": "As they have no fins, in case no fin is detected, the whole image is kept. Not exactly the perfect solution, but still a good simplification
", "votes": null, "replies": [] }, { "id": 1697800, "author_name": "remekkinas", "author_url": "", "post_date": "02/19/2022 21:05:35", "content": "Thank you for answer. Could you make screenshot of your background images? As I remember haar needs GT and background images. To be honest I always had problem with background which lead to poor result and this is why I do not use this kind of custom detector anymore.
", "votes": null, "replies": [] }, { "id": 1697810, "author_name": "jeanmilpied", "author_url": "", "post_date": "02/19/2022 21:14:29", "content": "For background images for training, I played silly:
\nA 100 images of lanscapes, a 100 scrapped images of sea and ocean and a 100 images of dogs
\nI used cats as landscapes but I thouth their ears really look like fins sometimes
OK. I see. Should work and it works. Great. I had always with this part problem. Thank you for advice.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1695993": "**haar cascade fins detector**\n> cleaning the image data is crucial for proper performance of classifiers! \n\nIn order to do so, please find shared in the link a custom haar-cascade filter specifically designed to detect fins of dolphins and whales in the image,\n\nhere is the link in my github: https://github.com/JeanMILPIED/happy_whale_and_dolphin_haar/blob/main/README.md\n\nyou can use the filter freely and add it in an image treatment loop,\n\nthis may sound very old school, but this is much more easy to compute than a yolovX deeplearning detection considering that almost all images have fins.\n\nfeel free to comment hereunder", "1696419": "Oh wow, thank you for sharing, @jeanmilpied! Amazing!", "1697635": "I pushed my f1 score for whale or dolphin prediction from 73% to 80% using the cropped images with my home made haar cascade fins detector. It is definitely not perfect but worth giving it a try", "1697685": "How have you managed to deal with bieluga and grey whale?", "1697797": "As they have no fins, in case no fin is detected, the whole image is kept. Not exactly the perfect solution, but still a good simplification", "1697800": "Thank you for answer. Could you make screenshot of your background images? As I remember haar needs GT and background images. To be honest I always had problem with background which lead to poor result and this is why I do not use this kind of custom detector anymore.", "1697810": "For background images for training, I played silly:\nA 100 images of lanscapes, a 100 scrapped images of sea and ocean and a 100 images of dogs\nI used cats as landscapes but I thouth their ears really look like fins sometimes", "1697814": "OK. I see. Should work and it works. Great. I had always with this part problem. Thank you for advice.", "1711898": "Dear Jean \nThank you for your comments. just a question. As I understood, you create your own dataset that contains images with fines, yes? what is your size of images?", "1712163": "True\nI made a folder with about 250 images of \"only\" fins based on the compet dataset" }, "source": "meta" }