\nOne easy thing is classification to species. The below figure is confusion matrix for my validation set.
\n
![\"newplot.png\"](\"https://i.postimg.cc/1zM2sSL3/newplot.png\")
I thank to organizers for this interesting challenge and congrats everyone who enjoyed it!
\nMy biggest mistake in this competition was my late participation. I wish I had joined sooner.
I would like to identify beluga. Let me write a little about why I thought this.
\nI don't have a much time to spend the competitions, so I was looking for the key point to boost the score. I was trying to figure out what was easy and what was difficult.
\nOne easy thing is classification to species. The below figure is confusion matrix for my validation set.
\n
It's clear that the almost every species can classify correctly for my validation set, though I have never tested test set.
\nNext I would like to find what species identify difficult.
\n![\"count-of-images.png\"](\"https://i.postimg.cc/Qd8vPVyy/count-of-images.png\")
\n![\"CV.png\"](\"https://i.postimg.cc/gjyBtPqp/CV.png\")
The first bar plot is number of images, and next one is CV for each species. I thought beluga could BOOST if this was improved because of the large number of images and low CV.
\nHowever I couldn't grasp good beluga feature by comparing wrong prediction images with correct images.
\nI visualized it with grad-cam because I thought the scar on the back might be a feature quantity. I thought it would have captured about the back.
![\"correct-0a380a45691ce5.jpg\"](\"https://i.postimg.cc/Gp2tcLHM/correct-0a380a45691ce5.jpg\")
![\"correct-12f80166d3f6e6.jpg\"](\"https://i.postimg.cc/CMD1M6xV/correct-12f80166d3f6e6.jpg\")
I only saw beluga because of lack of time, but I would have liked to have had a closer look at other species as well.
\nAnyway, I'm very looking forward to seeing the LB top solutions. I can't wait to see how they capture the whale's features.
Thank you!
", "messages": [ { "id": "1759821", "postDate": "04/19/2022 00:09:28", "content": "I thank to organizers for this interesting challenge and congrats everyone who enjoyed it!
\nMy biggest mistake in this competition was my late participation. I wish I had joined sooner.
I would like to identify beluga. Let me write a little about why I thought this.
\nI don't have a much time to spend the competitions, so I was looking for the key point to boost the score. I was trying to figure out what was easy and what was difficult.
\nOne easy thing is classification to species. The below figure is confusion matrix for my validation set.
\n
It's clear that the almost every species can classify correctly for my validation set, though I have never tested test set.
\nNext I would like to find what species identify difficult.
\n
\n
The first bar plot is number of images, and next one is CV for each species. I thought beluga could BOOST if this was improved because of the large number of images and low CV.
\nHowever I couldn't grasp good beluga feature by comparing wrong prediction images with correct images.
\nI visualized it with grad-cam because I thought the scar on the back might be a feature quantity. I thought it would have captured about the back.
I only saw beluga because of lack of time, but I would have liked to have had a closer look at other species as well.
\nAnyway, I'm very looking forward to seeing the LB top solutions. I can't wait to see how they capture the whale's features.
Thank you!
", "rawMarkdown": "I thank to organizers for this interesting challenge and congrats everyone who enjoyed it!\nMy biggest mistake in this competition was my late participation. I wish I had joined sooner.\n\nI would like to identify beluga. Let me write a little about why I thought this.\n\nI don't have a much time to spend the competitions, so I was looking for the key point to boost the score. I was trying to figure out what was easy and what was difficult.\nOne easy thing is classification to species. The below figure is confusion matrix for my validation set.\n[](https://postimg.cc/FfdDVMrM)\n\nIt's clear that the almost every species can classify correctly for my validation set, though I have never tested test set.\n\nNext I would like to find what species identify difficult.\n\n[](https://postimg.cc/D8NC4fFr)\n[](https://postimg.cc/YL93jZsD)\n\nThe first bar plot is number of images, and next one is CV for each species. I thought beluga could BOOST if this was improved because of the large number of images and low CV.\n\nHowever I couldn't grasp good beluga feature by comparing wrong prediction images with correct images. \nI visualized it with grad-cam because I thought the scar on the back might be a feature quantity. I thought it would have captured about the back.\n\n[](https://postimg.cc/0zTPnqGm)\n\n[](https://postimg.cc/8spDt4vy)\n\nI only saw beluga because of lack of time, but I would have liked to have had a closer look at other species as well.\nAnyway, I'm very looking forward to seeing the LB top solutions. I can't wait to see how they capture the whale's features.\n\nThank you!", "votes": null }, { "id": "1760427", "postDate": "04/19/2022 09:24:41", "content": "@gaborfodor - you are needed here ;)
", "rawMarkdown": "gaborfodor - you are needed here ;)", "votes": null }, { "id": "1760475", "postDate": "04/19/2022 10:14:03", "content": "Lol :D
\nTbh we did not do anything species dependent postprocessing. We trained custom backfin and full body detectors. Whenever we could not find a good confident backfin box we used full body or even original image as a fallback.
", "rawMarkdown": "Lol :D\n\nTbh we did not do anything species dependent postprocessing. We trained custom backfin and full body detectors. Whenever we could not find a good confident backfin box we used full body or even original image as a fallback.", "votes": null }, { "id": "1812548", "postDate": "06/06/2022 02:22:45", "content": "Thank you for your kindly comments! @gaborfodor
\nI figure out the solution!
@gaborfodor - you are needed here ;)
", "votes": null, "replies": [ { "id": 1760475, "author_name": "gaborfodor", "author_url": "", "post_date": "04/19/2022 10:14:03", "content": "Lol :D
\nTbh we did not do anything species dependent postprocessing. We trained custom backfin and full body detectors. Whenever we could not find a good confident backfin box we used full body or even original image as a fallback.
", "votes": null, "replies": [] }, { "id": 1812548, "author_name": "hattan0523", "author_url": "", "post_date": "06/06/2022 02:22:45", "content": "Thank you for your kindly comments! @gaborfodor
\nI figure out the solution!