|
Hi, I've had birdnet-go running for couple of months and everything is running smoothly. I really the service but sometimes I just don't understand the detection logic and was wondering if my settings are correct. I live in central Finland and here are couple of examples to discuss:
Is there any logic in birdnet which would prevent for example common cranes from being detected in the middle of winter? All the cranes should be in Africa. I have the range filter set up correctly but I guess it doesn't understand bird migrations / time of the year? I guess the crane is eurasian jay in my case but 100% common crane....? I really don't like having 100% certainty detections when they are not true in real life. What would be the reason for great tit not being detected even though they are chirping and singing multiple hours a day? What info is needed to dig deeper into this and is there something I can fine-tune.? Or have I found the limitations of the current capabilities? Thanks! |
Replies: 3 comments 2 replies
|
The only time I see truely unexpected birds, is when the wind is blowing hard. I would find it a disappointment if it only identified birds I expect. Everyday I see birds that are regional but never seen by me. It's why I have the software. I know how to ignore a false positive.
\\Greg
…________________________________
From: Antti ***@***.***>
Sent: Thursday, February 20, 2025 06:23
To: tphakala/birdnet-go ***@***.***>
Cc: Subscribed ***@***.***>
Subject: [tphakala/birdnet-go] Detections not being logical (Discussion #480)
Hi,
I've had birdnet-go running for couple of months and everything is running smoothly. I really the service but sometimes I just don't understand the detection logic and was wondering if my settings are correct. I live in central Finland and here are couple of examples to discuss:
1. I've gotten 100% common crane couple of times during winter (latest today)
2. I never have detections of great tit which is the most common bird on my yard
Is there any logic in birdnet which would prevent for example common cranes from being detected in the middle of winter? All the cranes should be in Africa. I have the range filter set up correctly but I guess it doesn't understand bird migrations / time of the year? I guess the crane is eurasian jay in my case but 100% common crane....? I really don't like having 100% certainty detections when they are not true in real life.
What would be the reason for great tit not being detected even though they are chirping and singing multiple hours a day? What info is needed to dig deeper into this and is there something I can fine-tune.? Or have I found the limitations of the current capabilities?
Thanks!
—
Reply to this email directly, view it on GitHub<#480>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AGEO6Z7BXCCDKA6MPYYAKE32QW3MHAVCNFSM6AAAAABXQQHTUCVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZXHE4TGOBYGU>.
You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>
|
|
Of course I'm also interested to see detections which are surprising and sometimes I get those as well. I would also like to gather some long term data to see how the amounts of birds change over time. Where I live it's actually very rare to get a bird which I haven't seen. I guess we have quite different environments! My point was to find out if there is something I need to tweak to make it even better. I really like the software like you do. |
|
Understandable. I believe You can block specific species on the species tab.
\\Greg
…________________________________
From: Antti ***@***.***>
Sent: Thursday, February 20, 2025 10:38:12 AM
To: tphakala/birdnet-go ***@***.***>
Cc: uSlackr ***@***.***>; Comment ***@***.***>
Subject: Re: [tphakala/birdnet-go] Detections not being logical (Discussion #480)
Of course I'm also interested to see detections which are surprising and sometimes I get those as well. I would also like to gather some long term data to see how the amounts of birds change over time. Where I live it's actually very rare to get a bird which I haven't seen. I guess we have quite different environments!
My point was to find out if there is something I need to tweak to make it even better. I really like the software like you do.
—
Reply to this email directly, view it on GitHub<#480 (reply in thread)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AGEO6ZZ3D5NILRGACHM42Z32QXZGJAVCNFSM6AAAAABXQQHTUCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTEMRWGUYTINA>.
You are receiving this because you commented.Message ID: ***@***.***>
|
|
@Tserppa The problem you are seeing is because the BirdNET AI model isn't perfect. Some species are clearly trained with "not so good" source data, and there is also some strong regional bias for some species due to birds having regional accents. I also have some crane matches which are clearly wrong at this time of year in Finland. These should be avoided by the range filter, which contains probabilities for every species based on date and location, but the range filter also seems to have imperfect training data. I am also experiencing very few detections for great tit, and have noticed similar issues with other local species. For example, Eurasian treecreeper is mostly detected as brown treecreeper (its American cousin), and the model has trouble with bullfinch. Also, when great tit calls are detected, they are often misclassified as Eurasian chaffinch. I have been training my own model to improve local species detections. I will make it available for everyone at some point, and also include it as the default model in BirdNET-Go, but this requires more improvements in both the model and BirdNET-Go side before I am willing to share it. I am trying to do my best to work around the limitations of BirdNET AI models by implementing some filters in BirdNET-Go. One is a time of day filter that I plan to implement at some point. If you wish to provide feedback about the BirdNET AI model, the best place to do this is at the BirdNET Analyzer repository discussion section: https://github.com/kahst/BirdNET-Analyzer/discussions |
|
Hi, thanks for your answer, this kind of information I was hoping to get! So basically we just don't have enough data from our birds/area to provide more accurate detections. My immediate though while reading your answer was that can we contribute to this project in some way but you have already started with your own model! I guess I could dive into the modeling rabbit hole myself... I will check the link as well! Kiitos työstäsi tämän eteen! |
@Tserppa The problem you are seeing is because the BirdNET AI model isn't perfect. Some species are clearly trained with "not so good" source data, and there is also some strong regional bias for some species due to birds having regional accents. I also have some crane matches which are clearly wrong at this time of year in Finland. These should be avoided by the range filter, which contains probabilities for every species based on date and location, but the range filter also seems to have imperfect training data.
I am also experiencing very few detections for great tit, and have noticed similar issues with other local species. For example, Eurasian treecreeper is mostly detected as brown…