This competitions is the multi labeling classifications.
\nNow, I have a question in this labels.
In the normal multi labeling classifications, I think that the label is lined up horizontally, for example, [1,4,5] etc…
\nHowever, the labels is grouped by the two classifications, primary_label and secondary_label in this competitions.
What is the meaning of dividing this?
\nMay I interplet that the bird sound of primary label is dominated?
This competitions is the multi labeling classifications.
\nNow, I have a question in this labels.
In the normal multi labeling classifications, I think that the label is lined up horizontally, for example, [1,4,5] etc…
\nHowever, the labels is grouped by the two classifications, primary_label and secondary_label in this competitions.
What is the meaning of dividing this?
\nMay I interplet that the bird sound of primary label is dominated?
That would probably be a logical assumption but far from being a rule. It's definitely a nice way to immediately bypass recordings of birds you don't want to potentially overlap without actually analysing the recordings autonomously or manually, since it's quite a narrow highway.
", "rawMarkdown": "That would probably be a logical assumption but far from being a rule. It's definitely a nice way to immediately bypass recordings of birds you don't want to potentially overlap without actually analysing the recordings autonomously or manually, since it's quite a narrow highway.", "votes": null }, { "id": "1293858", "postDate": "05/05/2021 08:05:33", "content": "Thank you for kindly description.
\nThat is, Doesn't it make sense to separate primary_label and secondary_label?
Been active on Kaggle for couple of years. A trend I think I see in recent competitions is a need by the hosts to have tools that handle weakly labeled data or very noisy data.
\nIt's very straight forward to create a working model that will give you decent accuracy if you train on sound clips that have high quality ratings and only a single primary bird call. Sadly the real world seldom provides high quality recordings with a single bird and the public/private test recordings are real world.
\nThe hosts have an existing app that I fired up two weeks ago to identify birds based on phone recordings. \"Birdnet\" is the Iphone app name. I take a daily walk of several miles along the Allegheny river and decided to try out the app. As I listened to the recordings I made I was amazed at how noisy the world around me was - having walked this same 4 mile route every day for most of the past decade all the \"noise\" had become white noise to me - be not to my phone recording. Multiple birds beyond the one of interest to me and lots of background noises in all the recordings I made.
\nI suggest loading up your I phone with the App - than attempting to get some bird calls. I think you will gain an appreciation that the hosts for this competition have a well working app when you can capture a recording with a single bird and limited background noises - they need our help with less than perfect recordings, which so far are 99% of the ones I have tried to capture.
\nA decent starting point would be to model high quality rated recordings with only single birds - but the model you generate will not likely get in the metal range for this competition and be of little assistance to the hosts. How to handle the secondary birds, low quality recordings and the \"noise\" is the real challenge in this competition.
", "rawMarkdown": "Been active on Kaggle for couple of years. A trend I think I see in recent competitions is a need by the hosts to have tools that handle weakly labeled data or very noisy data.\n\nIt's very straight forward to create a working model that will give you decent accuracy if you train on sound clips that have high quality ratings and only a single primary bird call. Sadly the real world seldom provides high quality recordings with a single bird and the public/private test recordings are real world.\n\nThe hosts have an existing app that I fired up two weeks ago to identify birds based on phone recordings. \"Birdnet\" is the Iphone app name. I take a daily walk of several miles along the Allegheny river and decided to try out the app. As I listened to the recordings I made I was amazed at how noisy the world around me was - having walked this same 4 mile route every day for most of the past decade all the \"noise\" had become white noise to me - be not to my phone recording. Multiple birds beyond the one of interest to me and lots of background noises in all the recordings I made. \n\nI suggest loading up your I phone with the App - than attempting to get some bird calls. I think you will gain an appreciation that the hosts for this competition have a well working app when you can capture a recording with a single bird and limited background noises - they need our help with less than perfect recordings, which so far are 99% of the ones I have tried to capture.\n\nA decent starting point would be to model high quality rated recordings with only single birds - but the model you generate will not likely get in the metal range for this competition and be of little assistance to the hosts. How to handle the secondary birds, low quality recordings and the \"noise\" is the real challenge in this competition.", "votes": null }, { "id": "1301992", "postDate": "05/11/2021 10:57:54", "content": "Thank you for kindly description. I clearly understood The meaning of label division.
\nI will try to use the \"Birdnet\".
That would probably be a logical assumption but far from being a rule. It's definitely a nice way to immediately bypass recordings of birds you don't want to potentially overlap without actually analysing the recordings autonomously or manually, since it's quite a narrow highway.
", "votes": null, "replies": [ { "id": 1293858, "author_name": "kunihikofurugori", "author_url": "", "post_date": "05/05/2021 08:05:33", "content": "Thank you for kindly description.
\nThat is, Doesn't it make sense to separate primary_label and secondary_label?
Been active on Kaggle for couple of years. A trend I think I see in recent competitions is a need by the hosts to have tools that handle weakly labeled data or very noisy data.
\nIt's very straight forward to create a working model that will give you decent accuracy if you train on sound clips that have high quality ratings and only a single primary bird call. Sadly the real world seldom provides high quality recordings with a single bird and the public/private test recordings are real world.
\nThe hosts have an existing app that I fired up two weeks ago to identify birds based on phone recordings. \"Birdnet\" is the Iphone app name. I take a daily walk of several miles along the Allegheny river and decided to try out the app. As I listened to the recordings I made I was amazed at how noisy the world around me was - having walked this same 4 mile route every day for most of the past decade all the \"noise\" had become white noise to me - be not to my phone recording. Multiple birds beyond the one of interest to me and lots of background noises in all the recordings I made.
\nI suggest loading up your I phone with the App - than attempting to get some bird calls. I think you will gain an appreciation that the hosts for this competition have a well working app when you can capture a recording with a single bird and limited background noises - they need our help with less than perfect recordings, which so far are 99% of the ones I have tried to capture.
\nA decent starting point would be to model high quality rated recordings with only single birds - but the model you generate will not likely get in the metal range for this competition and be of little assistance to the hosts. How to handle the secondary birds, low quality recordings and the \"noise\" is the real challenge in this competition.
", "votes": null, "replies": [ { "id": 1301992, "author_name": "kunihikofurugori", "author_url": "", "post_date": "05/11/2021 10:57:54", "content": "Thank you for kindly description. I clearly understood The meaning of label division.
\nI will try to use the \"Birdnet\".