{ "id": 308573, "title": "Starting with domain knowledge and FIN-PRINT - inspiration how to approach problem", "url": "/competitions/happy-whale-and-dolphin/discussion/308573", "author_name": "", "post_date": "2022-02-19T09:29:02.708846400Z", "votes": 27, "comment_count": 12, "views": 0, "content": "

After TF competition I switched here and first step was to understand domain knowledge of competition.

\n

Source: https://wildwhales.org/speciesid/

\n\n

The most interesting part from competition perspective is certainly Common Dorsal Fins chapters which describe characteristic of Fins.

\n

Conclutions:

\n\n

Possible solutions:

\n\n

Next I found: https://www.nature.com/articles/s41598-021-02506-6 FIN-PRINT a fully-automated multi-stage deep-learning-based framework for the individual recognition of killer whales. Which can be really inspiring. Authors describe way to identify whales based on Dorsal Fins.
\nThe github repository you can find here https://github.com/ChristianBergler . A lot of tools for us in this competion.

\n

Full proposed pipeline looks like:

\n

\"Pipeline\"

\n

Quick summary:
\nThe FIN-PRINT pipeline consists of FIN-DETECT, a YOLOv3 -based object detection network for recognizing killer whale dorsal fins and associated saddle patches in images with 1 to N individuals, FIN-EXTRACT, an automatic extraction procedure cropping and equally resizing all detected dorsal fin/saddle patch markings within an image, VVI-DETECT, a ResNet34-based convolutional neural network (CNN) performing data enhancement by classifying between previously detected/extracted valid versus invalid (VVI) killer whale identification sub-images (e.g. bad weather conditions, blurred, missing saddle patch, difficult angle, detection errors, etc.), and FIN-IDENTIFY, a ResNet34-based CNN for multi-class killer whale individual classification modeling the 100 most commonly photo-identified killer whales. To the best of the authors’ knowledge, this is the first study transferring the analysis of killer whale image identification into a fully automated, multi-stage, sequentially ordered, deep-learning-based framework, in order to machine-identify individuals. Source: https://www.nature.com/articles/s41598-021-02506-6

\n

I thnk that it is really good approach which adresses most of problems described in post created by @andradaolteanu Things to know before starting image preprocessing which is BTW really great (it shows problematic group very well).

", "messages": [ { "id": "1696980", "postDate": "02/19/2022 09:29:02", "content": "

After TF competition I switched here and first step was to understand domain knowledge of competition.

\n

Source: https://wildwhales.org/speciesid/

\n\n

The most interesting part from competition perspective is certainly Common Dorsal Fins chapters which describe characteristic of Fins.

\n

Conclutions:

\n\n

Possible solutions:

\n\n

Next I found: https://www.nature.com/articles/s41598-021-02506-6 FIN-PRINT a fully-automated multi-stage deep-learning-based framework for the individual recognition of killer whales. Which can be really inspiring. Authors describe way to identify whales based on Dorsal Fins.
\nThe github repository you can find here https://github.com/ChristianBergler . A lot of tools for us in this competion.

\n

Full proposed pipeline looks like:

\n

\"Pipeline\"

\n

Quick summary:
\nThe FIN-PRINT pipeline consists of FIN-DETECT, a YOLOv3 -based object detection network for recognizing killer whale dorsal fins and associated saddle patches in images with 1 to N individuals, FIN-EXTRACT, an automatic extraction procedure cropping and equally resizing all detected dorsal fin/saddle patch markings within an image, VVI-DETECT, a ResNet34-based convolutional neural network (CNN) performing data enhancement by classifying between previously detected/extracted valid versus invalid (VVI) killer whale identification sub-images (e.g. bad weather conditions, blurred, missing saddle patch, difficult angle, detection errors, etc.), and FIN-IDENTIFY, a ResNet34-based CNN for multi-class killer whale individual classification modeling the 100 most commonly photo-identified killer whales. To the best of the authors’ knowledge, this is the first study transferring the analysis of killer whale image identification into a fully automated, multi-stage, sequentially ordered, deep-learning-based framework, in order to machine-identify individuals. Source: https://www.nature.com/articles/s41598-021-02506-6

\n

I thnk that it is really good approach which adresses most of problems described in post created by @andradaolteanu Things to know before starting image preprocessing which is BTW really great (it shows problematic group very well).

", "rawMarkdown": "After TF competition I switched here and first step was to understand domain knowledge of competition. \n\nSource: https://wildwhales.org/speciesid/\n- What is that whale?\n- Characteristic of Common Dorsal Fins\n- Common Blows\n- Common Flukes\n- Common Behaviours\n- Common Splashes\n\nThe most interesting part from competition perspective is certainly Common Dorsal Fins chapters which describe characteristic of Fins. \n\nConclutions:\n- we can identify species by the **characteristics, shape, size (1)** and **position of the fin (2)** eg. Killer Whale - curved towards the back, up to 1.8 metres in height, in the middle of the back\n- there are spcies without dorsal fin - bieluga and grey whale (no dorsal fin, only a small hump followed by a series of ‘knuckles’ 2/3 of the way back from the head)\n\nPossible solutions:\n- **check characteristics (1)** - feature extraction (patterns etc.)\n- **locate fin, locate whale (2)** and check relationship - could be hard in most cases since we have only dorsal fin on photo (but possible)\n\nNext I found: https://www.nature.com/articles/s41598-021-02506-6 **FIN-PRINT a fully-automated multi-stage deep-learning-based framework for the individual recognition of killer whales**. Which can be really inspiring. Authors describe way to identify whales based on Dorsal Fins.\nThe github repository you can find here https://github.com/ChristianBergler . A lot of tools for us in this competion. \n\nFull proposed pipeline looks like:\n\n![Pipeline](https://i.ibb.co/vvVCCPr/41598-2021-2506-Fig1-HTML.webp)\n\nQuick summary:\nThe FIN-PRINT pipeline consists of FIN-DETECT, a YOLOv3 -based object detection network for recognizing killer whale dorsal fins and associated saddle patches in images with 1 to N individuals, FIN-EXTRACT, an automatic extraction procedure cropping and equally resizing all detected dorsal fin/saddle patch markings within an image, VVI-DETECT, a ResNet34-based convolutional neural network (CNN) performing data enhancement by classifying between previously detected/extracted valid versus invalid (VVI) killer whale identification sub-images (e.g. bad weather conditions, blurred, missing saddle patch, difficult angle, detection errors, etc.), and FIN-IDENTIFY, a ResNet34-based CNN for multi-class killer whale individual classification modeling the 100 most commonly photo-identified killer whales. To the best of the authors’ knowledge, this is the first study transferring the analysis of killer whale image identification into a fully automated, multi-stage, sequentially ordered, deep-learning-based framework, in order to machine-identify individuals. Source: https://www.nature.com/articles/s41598-021-02506-6\n\nI thnk that it is really good approach which adresses most of problems described in post created by @andradaolteanu [Things to know before starting image preprocessing](https://www.kaggle.com/c/happy-whale-and-dolphin/discussion/308026) which is BTW really great (it shows problematic group very well).", "votes": null }, { "id": "1697081", "postDate": "02/19/2022 11:06:01", "content": "

Nice to see you here @remekkinas , your contributions are always helpful

", "rawMarkdown": "Nice to see you here @remekkinas , your contributions are always helpful", "votes": null }, { "id": "1697088", "postDate": "02/19/2022 11:11:55", "content": "

Thank you very much :) I had to rest a little bit after TF :) I was really exhausted 😤😤😳

\n

I thnink that FIN-PRINT is really good way to follow. First deal with data and create data more understandable to model. BTW: nice competition.

", "rawMarkdown": "Thank you very much :) I had to rest a little bit after TF :) I was really exhausted 😤😤😳\n\nI thnink that FIN-PRINT is really good way to follow. First deal with data and create data more understandable to model. BTW: nice competition.", "votes": null }, { "id": "1697092", "postDate": "02/19/2022 11:16:04", "content": "
\n

I had to res a little bit after TF :)

\n
\n

Less than a week and you're back already? 🙏😄

", "rawMarkdown": "> I had to res a little bit after TF :)\n\nLess than a week and you're back already? 🙏😄", "votes": null }, { "id": "1697337", "postDate": "02/19/2022 14:52:03", "content": "

This post is amazing and answers a few of the questions that I had on how to approach this dataset. Thank you sososo much for the info!

\n

Also, I am very glad you found my post helpful 🙏

", "rawMarkdown": "This post is amazing and answers a few of the questions that I had on how to approach this dataset. Thank you sososo much for the info!\n\nAlso, I am very glad you found my post helpful 🙏", "votes": null }, { "id": "1697354", "postDate": "02/19/2022 15:04:42", "content": "

Yes, absolutely your post is outstending. You saved a lot of my time. I downloaded data and was categorizing and looking for problematic group. Then found your post and … I was impressed. Absolutely good approach to computer vision task. Look into data, divide into problematic group and propose solution for each of group.

\n

I agree FIN-PRINT adresses many problems we have her :)

", "rawMarkdown": "Yes, absolutely your post is outstending. You saved a lot of my time. I downloaded data and was categorizing and looking for problematic group. Then found your post and ... I was impressed. Absolutely good approach to computer vision task. Look into data, divide into problematic group and propose solution for each of group. \n\nI agree FIN-PRINT adresses many problems we have her :)", "votes": null }, { "id": "1697556", "postDate": "02/19/2022 17:27:37", "content": "

Great insights, you answered some questions! Good luck, @remekkinas!

", "rawMarkdown": "Great insights, you answered some questions! Good luck, @remekkinas!", "votes": null }, { "id": "1697718", "postDate": "02/19/2022 19:50:19", "content": "

You are welcome. First stages in proposed attitude are probably the most important in this competition. We will see it soon.

", "rawMarkdown": "You are welcome. First stages in proposed attitude are probably the most important in this competition. We will see it soon.", "votes": null }, { "id": "1698316", "postDate": "02/20/2022 09:43:00", "content": "

Yes, we do not want to lose momentum ;) I am looking for first gold in competition 😂😂😃😁

", "rawMarkdown": "Yes, we do not want to lose momentum ;) I am looking for first gold in competition 😂😂😃😁", "votes": null }, { "id": "1698602", "postDate": "02/20/2022 14:19:01", "content": "

Great notes. Check this out now!

", "rawMarkdown": "Great notes. Check this out now!", "votes": null }, { "id": "1698618", "postDate": "02/20/2022 14:33:06", "content": "

Good Luck @remekkinas, I've just come out of my \"Kaggle chai break\" but I have enjoyed interacting with you on the forums. I hope you become a Comp master soon! :)

", "rawMarkdown": "Good Luck @remekkinas, I've just come out of my \"Kaggle chai break\" but I have enjoyed interacting with you on the forums. I hope you become a Comp master soon! :)", "votes": null }, { "id": "1698671", "postDate": "02/20/2022 15:09:35", "content": "

Really helpful note. It's cheering that you are here.

", "rawMarkdown": "Really helpful note. It's cheering that you are here.", "votes": null }, { "id": "1698722", "postDate": "02/20/2022 15:57:41", "content": "

Amazing, nice to see you again!

", "rawMarkdown": "Amazing, nice to see you again!", "votes": null } ], "comments": [ { "id": 1697081, "author_name": "mrinath", "author_url": "", "post_date": "02/19/2022 11:06:01", "content": "

Nice to see you here @remekkinas , your contributions are always helpful

", "votes": null, "replies": [ { "id": 1697088, "author_name": "remekkinas", "author_url": "", "post_date": "02/19/2022 11:11:55", "content": "

Thank you very much :) I had to rest a little bit after TF :) I was really exhausted 😤😤😳

\n

I thnink that FIN-PRINT is really good way to follow. First deal with data and create data more understandable to model. BTW: nice competition.

", "votes": null, "replies": [] }, { "id": 1697092, "author_name": "init27", "author_url": "", "post_date": "02/19/2022 11:16:04", "content": "
\n

I had to res a little bit after TF :)

\n
\n

Less than a week and you're back already? 🙏😄

", "votes": null, "replies": [] }, { "id": 1698316, "author_name": "remekkinas", "author_url": "", "post_date": "02/20/2022 09:43:00", "content": "

Yes, we do not want to lose momentum ;) I am looking for first gold in competition 😂😂😃😁

", "votes": null, "replies": [] }, { "id": 1698618, "author_name": "init27", "author_url": "", "post_date": "02/20/2022 14:33:06", "content": "

Good Luck @remekkinas, I've just come out of my \"Kaggle chai break\" but I have enjoyed interacting with you on the forums. I hope you become a Comp master soon! :)

", "votes": null, "replies": [] } ] }, { "id": 1697337, "author_name": "andradaolteanu", "author_url": "", "post_date": "02/19/2022 14:52:03", "content": "

This post is amazing and answers a few of the questions that I had on how to approach this dataset. Thank you sososo much for the info!

\n

Also, I am very glad you found my post helpful 🙏

", "votes": null, "replies": [ { "id": 1697354, "author_name": "remekkinas", "author_url": "", "post_date": "02/19/2022 15:04:42", "content": "

Yes, absolutely your post is outstending. You saved a lot of my time. I downloaded data and was categorizing and looking for problematic group. Then found your post and … I was impressed. Absolutely good approach to computer vision task. Look into data, divide into problematic group and propose solution for each of group.

\n

I agree FIN-PRINT adresses many problems we have her :)

", "votes": null, "replies": [] } ] }, { "id": 1697556, "author_name": "vad13irt", "author_url": "", "post_date": "02/19/2022 17:27:37", "content": "

Great insights, you answered some questions! Good luck, @remekkinas!

", "votes": null, "replies": [ { "id": 1697718, "author_name": "remekkinas", "author_url": "", "post_date": "02/19/2022 19:50:19", "content": "

You are welcome. First stages in proposed attitude are probably the most important in this competition. We will see it soon.

", "votes": null, "replies": [] } ] }, { "id": 1698602, "author_name": "crained", "author_url": "", "post_date": "02/20/2022 14:19:01", "content": "

Great notes. Check this out now!

", "votes": null, "replies": [] }, { "id": 1698671, "author_name": "snaker", "author_url": "", "post_date": "02/20/2022 15:09:35", "content": "

Really helpful note. It's cheering that you are here.

", "votes": null, "replies": [] }, { "id": 1698722, "author_name": "vincentwang25", "author_url": "", "post_date": "02/20/2022 15:57:41", "content": "

Amazing, nice to see you again!

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1696980": "After TF competition I switched here and first step was to understand domain knowledge of competition. \n\nSource: https://wildwhales.org/speciesid/\n- What is that whale?\n- Characteristic of Common Dorsal Fins\n- Common Blows\n- Common Flukes\n- Common Behaviours\n- Common Splashes\n\nThe most interesting part from competition perspective is certainly Common Dorsal Fins chapters which describe characteristic of Fins. \n\nConclutions:\n- we can identify species by the **characteristics, shape, size (1)** and **position of the fin (2)** eg. Killer Whale - curved towards the back, up to 1.8 metres in height, in the middle of the back\n- there are spcies without dorsal fin - bieluga and grey whale (no dorsal fin, only a small hump followed by a series of ‘knuckles’ 2/3 of the way back from the head)\n\nPossible solutions:\n- **check characteristics (1)** - feature extraction (patterns etc.)\n- **locate fin, locate whale (2)** and check relationship - could be hard in most cases since we have only dorsal fin on photo (but possible)\n\nNext I found: https://www.nature.com/articles/s41598-021-02506-6 **FIN-PRINT a fully-automated multi-stage deep-learning-based framework for the individual recognition of killer whales**. Which can be really inspiring. Authors describe way to identify whales based on Dorsal Fins.\nThe github repository you can find here https://github.com/ChristianBergler . A lot of tools for us in this competion. \n\nFull proposed pipeline looks like:\n\n![Pipeline](https://i.ibb.co/vvVCCPr/41598-2021-2506-Fig1-HTML.webp)\n\nQuick summary:\nThe FIN-PRINT pipeline consists of FIN-DETECT, a YOLOv3 -based object detection network for recognizing killer whale dorsal fins and associated saddle patches in images with 1 to N individuals, FIN-EXTRACT, an automatic extraction procedure cropping and equally resizing all detected dorsal fin/saddle patch markings within an image, VVI-DETECT, a ResNet34-based convolutional neural network (CNN) performing data enhancement by classifying between previously detected/extracted valid versus invalid (VVI) killer whale identification sub-images (e.g. bad weather conditions, blurred, missing saddle patch, difficult angle, detection errors, etc.), and FIN-IDENTIFY, a ResNet34-based CNN for multi-class killer whale individual classification modeling the 100 most commonly photo-identified killer whales. To the best of the authors’ knowledge, this is the first study transferring the analysis of killer whale image identification into a fully automated, multi-stage, sequentially ordered, deep-learning-based framework, in order to machine-identify individuals. Source: https://www.nature.com/articles/s41598-021-02506-6\n\nI thnk that it is really good approach which adresses most of problems described in post created by @andradaolteanu [Things to know before starting image preprocessing](https://www.kaggle.com/c/happy-whale-and-dolphin/discussion/308026) which is BTW really great (it shows problematic group very well).", "1697081": "Nice to see you here @remekkinas , your contributions are always helpful", "1697088": "Thank you very much :) I had to rest a little bit after TF :) I was really exhausted 😤😤😳\n\nI thnink that FIN-PRINT is really good way to follow. First deal with data and create data more understandable to model. BTW: nice competition.", "1697092": "> I had to res a little bit after TF :)\n\nLess than a week and you're back already? 🙏😄", "1697337": "This post is amazing and answers a few of the questions that I had on how to approach this dataset. Thank you sososo much for the info!\n\nAlso, I am very glad you found my post helpful 🙏", "1697354": "Yes, absolutely your post is outstending. You saved a lot of my time. I downloaded data and was categorizing and looking for problematic group. Then found your post and ... I was impressed. Absolutely good approach to computer vision task. Look into data, divide into problematic group and propose solution for each of group. \n\nI agree FIN-PRINT adresses many problems we have her :)", "1697556": "Great insights, you answered some questions! Good luck, @remekkinas!", "1697718": "You are welcome. First stages in proposed attitude are probably the most important in this competition. We will see it soon.", "1698316": "Yes, we do not want to lose momentum ;) I am looking for first gold in competition 😂😂😃😁", "1698602": "Great notes. Check this out now!", "1698618": "Good Luck @remekkinas, I've just come out of my \"Kaggle chai break\" but I have enjoyed interacting with you on the forums. I hope you become a Comp master soon! :)", "1698671": "Really helpful note. It's cheering that you are here.", "1698722": "Amazing, nice to see you again!" }, "source": "meta" }