search

ilterturkmenli / HistSegNet

1 stars 0 forks source link

Hello，I have a problems about sen1Floods11 #1

Open sam-sun-2001 opened 2 weeks ago

sam-sun-2001 commented 2 weeks ago

when I use the dataset to train, I only used all SAR data from the flood event. the results showed me the miou is 75%. when I look the related papar, I found most of them showed the iou is only 40%. Why so great difference ? Could you tell me ?

ilterturkmenli commented 2 weeks ago

The reason of that probably you calculate the miou over an confusion matrix. You should first calculate iou for each 512x512 patch, then take the mean of them. And be carefull to pixel label 255, you should not account them in evaluation.

Android için Outlookhttps://aka.ms/AAb9ysg edinin

From: sam-sun-2001 @.> Sent: Saturday, November 16, 2024 10:33:55 AM To: ilterturkmenli/HistSegNet @.> Cc: Subscribed @.***> Subject: [ilterturkmenli/HistSegNet] Hello，I have a problems about sen1Floods11 (Issue #1)

when I use the dataset to train, I only used all SAR data from the flood event. the results showed me the miou is 75%. when I look the related papar, I found most of them showed the iou is only 40%. Why so great difference ? Could you tell me ?

— Reply to this email directly, view it on GitHubhttps://github.com/ilterturkmenli/HistSegNet/issues/1, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AGVJTDCIQ4N7XNYSCSNIX3L2A3YOHAVCNFSM6AAAAABR4RXY42VHI2DSMVQWIX3LMV43ASLTON2WKOZSGY3DGOJXG4YDSMI. You are receiving this because you are subscribed to this thread.Message ID: @.***>

sam-sun-2001 commented 2 days ago

Thank you so much for your reply. I appreciate it. And I would like to ask a question, Do I need to focus miou or just focus on the iou for flood? and I train data then pick the hightest iou validation pt to test test_dataset, is my logic correct ? The last question is that do you know any sota model on the sen1floods11 dataset except BAS model? Thank you so much ! I changed my account so I saw you reply just now.

sunhanzhang 201918020233

On 2024-11-16 17:40, ilterturkmenli wrote:

The reason of that probably you calculate the miou over an confusion matrix. You should first calculate iou for each 512x512 patch, then take the mean of them. And be carefull to pixel label 255, you should not account them in evaluation.

Android için Outlookhttps://aka.ms/AAb9ysg edinin

From: sam-sun-2001 @.> Sent: Saturday, November 16, 2024 10:33:55 AM To: ilterturkmenli/HistSegNet @.> Cc: Subscribed @.***> Subject: [ilterturkmenli/HistSegNet] Hello，I have a problems about sen1Floods11 (Issue #1)

when I use the dataset to train, I only used all SAR data from the flood event. the results showed me the miou is 75%. when I look the related papar, I found most of them showed the iou is only 40%. Why so great difference ? Could you tell me ?

-- Reply to this email directly, view it on GitHubhttps://github.com/ilterturkmenli/HistSegNet/issues/1, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AGVJTDCIQ4N7XNYSCSNIX3L2A3YOHAVCNFSM6AAAAABR4RXY42VHI2DSMVQWIX3LMV43ASLTON2WKOZSGY3DGOJXG4YDSMI. You are receiving this because you are subscribed to this thread.Message ID: @.***>

Reply to this email directly, view it on GitHub [1], or unsubscribe [2]. You are receiving this because you authored the thread.Message ID: @.***>

Links:

[1] https://github.com/ilterturkmenli/HistSegNet/issues/1#issuecomment-2480495241 [2] https://github.com/notifications/unsubscribe-auth/AWOKNTET6JF3NIIQUZVBZYL2A4HITAVCNFSM6AAAAABR4RXY42VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIOBQGQ4TKMRUGE --=_03d195b414aecbb9afb8d427c28b6116 Content-Type: multipart/related; boundary="=_382f1f6e1b702c2676b97739e37bc57d"

--=_382f1f6e1b702c2676b97739e37bc57d Content-Transfer-Encoding: quoted-printable Content-Type: text/html; charset=UTF-8

Thank you so much for your reply. I appreciate it. And I would like to a= sk a question, Do I need to focus miou or just focus on the iou for flood?&= nbsp;  and I train data then pick the hightest iou validation pt to te= st test_dataset, is my logic correct ? The last question is that do you kno= w any sota model on the sen1floods11 dataset except BAS model?
Thank y= ou so much !  I changed my account so I saw you reply just now.

---
= sunhanzhang  201918020233


On 2024-11-16 17:40, ilterturkmenli wrote:


The reason of that probably you calculate the miou over an confusion matrix= =2E You should first calculate iou for each 512x512 patch, then take the me= an of them. And be carefull to pixel label 255, you should not account them= in evaluation.

Android için Outlook<https://aka.ms/A= Ab9ysg> edinin
________________________________
From: sam-su= n-2001 ***@***.***>
Sent: Saturday, November 16, 2024 10:33:55 AM =
To: ilterturkmenli/HistSegNet ***@***.***>
Cc: Subscribed **= *@***.***>
Subject: [ilterturkmenli/HistSegNet] Hello=EF=BC=8CI ha= ve a problems about sen1Floods11 (Issue #1)


when I use th= e dataset to train, I only used all SAR data from the flood event.
th= e results showed me the miou is 75%. when I look the related papar, I found= most of them showed the iou is only 40%. Why so great difference ? Could y= ou tell me ?

—
Reply to this email directly, view i= t on GitHub<https://github.com/ilterturkmenli/HistSegNet/issues/1>, o= r unsubscribe<https://github.com/notifications/unsubscribe-auth/AGVJTDCI= Q4N7XNYSCSNIX3L2A3YOHAVCNFSM6AAAAABR4RXY42VHI2DSMVQWIX3LMV43ASLTON2WKOZSGY3= DGOJXG4YDSMI>.
You are receiving this because you are subscribed t= o this thread.Message ID: ***@***.***>

—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are= receiving this because you authored the thread.Message ID: <ilterturkmenli/HistSe= gNet/issues/1/2480495241@github.com>

--=_382f1f6e1b702c2676b97739e37bc57d Content-Transfer-Encoding: base64 Content-ID: @.***> Content-Type: image/gif; name=blocked.gif Content-Disposition: inline; filename=blocked.gif; size=118

R0lGODlhZAAyAIAAAPrOzgAAACH5BAAAAAAALAAAAABkADIAAAJNhI+py+0Po5y02ouz3rz7D4bi SJbmiabqyrbuC8fyTNf2jef6zvf+DwwKh8Si8YhMKpfMpvMJjUqn1Kr1is1qt9yu9wsOi8fksvls KwAAOw== --=_382f1f6e1b702c2676b97739e37bc57d--

--=_03d195b414aecbb9afb8d427c28b6116--

sam-sun-2001 commented 2 days ago

Yes, you're right. I used a global confusion matrix to calculate both the mIoU and the IoU for the positive class. Isn't this a common approach? Can I continue using this method?