Input image pre-processing to avoid missed checks (allowing for multi-marked significant detection)?

[tongxiaozhong14]

Dear author, there are still some missed checks when testing this code? How should I pre-process the data set (e.g. brighten or darken?) to ensure that no inspection is missed (even a few more notable marks are fine)?

[xuebinqin]

Thank you for your interests. But I didn't get your point. I am sorry about that. What do you mean about "some missed checks"? As for the pre-processing, you can check the dataloader.py. We only conduct geometric transforman and simple normalization of the input images. No more intensity-based transformations are used.

On Sun, May 17, 2020 at 6:51 AM tongxiaozhong14 notifications@github.com wrote:

Dear author, there are still some missed checks when testing this code? How should I pre-process the data set (e.g. brighten or darken?) to ensure that no inspection is missed (even a few more notable marks are fine)?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/NathanUA/BASNet/issues/38, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORPFRWQB43T4NDR3IL3RR7MULANCNFSM4NDL3GVA .

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/

[tongxiaozhong14]

Dear author, Here's the thing, I took your code and tested my own dataset, but there are some significant target omissions that occur. The questions I would like to ask you are: (1) How do I handle (e.g., brighten or darken, etc.) my images to ensure that your BASNet.py code does not leak? (It doesn't matter even if a few more significant targets are detected,) Most importantly, just make sure there are no missed inspections, no missed inspections allowed which is important to me!!!!

[tongxiaozhong14]

Thank you for your interests. But I didn't get your point. I am sorry about that. What do you mean about "some missed checks"? As for the pre-processing, you can check the dataloader.py. We only conduct geometric transforman and simple normalization of the input images. No more intensity-based transformations are used. … On Sun, May 17, 2020 at 6:51 AM tongxiaozhong14 @.***> wrote: Dear author, there are still some missed checks when testing this code? How should I pre-process the data set (e.g. brighten or darken?) to ensure that no inspection is missed (even a few more notable marks are fine)? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#38>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORPFRWQB43T4NDR3IL3RR7MULANCNFSM4NDL3GVA . -- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/

Dear author, Here's the thing, I took your code and tested my own dataset, but there are some significant target omissions that occur. The questions I would like to ask you are: (1) How do I handle (e.g., brighten or darken, etc.) my images to ensure that your BASNet.py code does not leak? (It doesn't matter even if a few more significant targets are detected,) Most importantly, just make sure there are no missed inspections, no missed inspections allowed which is important to me!!!!

[xuebinqin]

I am not sure what kind of dataset are you using. The network usually not only learns the image contrast but also fits the shape and position of the targets. So sometimes, even the target is very salient as we human see, but the network may fail because there were no similar images in the training set. Therefore, to get good results on your dataset, the best way is retrain ingthe network on your dataset. But it takes a lot fo resources for labeling images. There still some methods you can try but the results are not guaranteed. (1) try to equalize the image intensity to increase the image contrast (https://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_equalize.html https://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_equalize.html), (2) enlarge the image with blank background ( changes the relative position and scale of the targets and the whole image).

On May 19, 2020, at 5:57 AM, tongxiaozhong14 notifications@github.com wrote:

Thank you for your interests. But I didn't get your point. I am sorry about that. What do you mean about "some missed checks"? As for the pre-processing, you can check the dataloader.py. We only conduct geometric transforman and simple normalization of the input images. No more intensity-based transformations are used. … <x-msg://22/#> On Sun, May 17, 2020 at 6:51 AM tongxiaozhong14 @.***> wrote: Dear author, there are still some missed checks when testing this code? How should I pre-process the data set (e.g. brighten or darken?) to ensure that no inspection is missed (even a few more notable marks are fine)? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#38 https://github.com/NathanUA/BASNet/issues/38>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORPFRWQB43T4NDR3IL3RR7MULANCNFSM4NDL3GVA https://github.com/notifications/unsubscribe-auth/ADSGORPFRWQB43T4NDR3IL3RR7MULANCNFSM4NDL3GVA . -- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/ https://webdocs.cs.ualberta.ca/~xuebin/ Dear author, Here's the thing, I took your code and tested my own dataset, but there are some significant target omissions that occur. The questions I would like to ask you are: (1) How do I handle (e.g., brighten or darken, etc.) my images to ensure that your BASNet.py code does not leak? (It doesn't matter even if a few more significant targets are detected,) Most importantly, just make sure there are no missed inspections, no missed inspections allowed which is important to me!!!!

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/NathanUA/BASNet/issues/38#issuecomment-630771144, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORKIZTHLV7NAPOQ5OSTRSJX3FANCNFSM4NDL3GVA.

[tongxiaozhong14]

I am not sure what kind of dataset are you using. The network usually not only learns the image contrast but also fits the shape and position of the targets. So sometimes, even the target is very salient as we human see, but the network may fail because there were no similar images in the training set. Therefore, to get good results on your dataset, the best way is retrain ingthe network on your dataset. But it takes a lot fo resources for labeling images. There still some methods you can try but the results are not guaranteed. (1) try to equalize the image intensity to increase the image contrast (https://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_equalize.html https://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_equalize.html), (2) enlarge the image with blank background ( changes the relative position and scale of the targets and the whole image). … On May 19, 2020, at 5:57 AM, tongxiaozhong14 @.> wrote: Thank you for your interests. But I didn't get your point. I am sorry about that. What do you mean about "some missed checks"? As for the pre-processing, you can check the dataloader.py. We only conduct geometric transforman and simple normalization of the input images. No more intensity-based transformations are used. … <x-msg://22/#> On Sun, May 17, 2020 at 6:51 AM tongxiaozhong14 @.> wrote: Dear author, there are still some missed checks when testing this code? How should I pre-process the data set (e.g. brighten or darken?) to ensure that no inspection is missed (even a few more notable marks are fine)? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#38 <#38>>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORPFRWQB43T4NDR3IL3RR7MULANCNFSM4NDL3GVA https://github.com/notifications/unsubscribe-auth/ADSGORPFRWQB43T4NDR3IL3RR7MULANCNFSM4NDL3GVA . -- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/ https://webdocs.cs.ualberta.ca/~xuebin/ Dear author, Here's the thing, I took your code and tested my own dataset, but there are some significant target omissions that occur. The questions I would like to ask you are: (1) How do I handle (e.g., brighten or darken, etc.) my images to ensure that your BASNet.py code does not leak? (It doesn't matter even if a few more significant targets are detected,) Most importantly, just make sure there are no missed inspections, no missed inspections allowed which is important to me!!!! — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#38 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORKIZTHLV7NAPOQ5OSTRSJX3FANCNFSM4NDL3GVA.

OK, thank you for your patient and detailed answers. Thank you from the bottom of my heart.

[xuebinqin]

You're very welcome.

On Thu, May 21, 2020 at 6:47 PM tongxiaozhong14 notifications@github.com wrote:

I am not sure what kind of dataset are you using. The network usually not only learns the image contrast but also fits the shape and position of the targets. So sometimes, even the target is very salient as we human see, but the network may fail because there were no similar images in the training set. Therefore, to get good results on your dataset, the best way is retrain ingthe network on your dataset. But it takes a lot fo resources for labeling images. There still some methods you can try but the results are not guaranteed. (1) try to equalize the image intensity to increase the image contrast ( https://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_equalize.html https://scikit-image.org/docs/dev/auto_examples/color_exposure/plot_equalize.html), (2) enlarge the image with blank background ( changes the relative position and scale of the targets and the whole image). … <#m-1333911955135443944> On May 19, 2020, at 5:57 AM, tongxiaozhong14 @.> wrote: Thank you for your interests. But I didn't get your point. I am sorry about that. What do you mean about "some missed checks"? As for the pre-processing, you can check the dataloader.py. We only conduct geometric transforman and simple normalization of the input images. No more intensity-based transformations are used. … x-msg://22/# On Sun, May 17, 2020 at 6:51 AM tongxiaozhong14 @.> wrote: Dear author, there are still some missed checks when testing this code? How should I pre-process the data set (e.g. brighten or darken?) to ensure that no inspection is missed (even a few more notable marks are fine)? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#38 https://github.com/NathanUA/BASNet/issues/38 <#38 https://github.com/NathanUA/BASNet/issues/38>>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORPFRWQB43T4NDR3IL3RR7MULANCNFSM4NDL3GVA https://github.com/notifications/unsubscribe-auth/ADSGORPFRWQB43T4NDR3IL3RR7MULANCNFSM4NDL3GVA . -- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/ https://webdocs.cs.ualberta.ca/~xuebin/ Dear author, Here's the thing, I took your code and tested my own dataset, but there are some significant target omissions that occur. The questions I would like to ask you are: (1) How do I handle (e.g., brighten or darken, etc.) my images to ensure that your BASNet.py code does not leak? (It doesn't matter even if a few more significant targets are detected,) Most importantly, just make sure there are no missed inspections, no missed inspections allowed which is important to me!!!! — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#38 (comment) https://github.com/NathanUA/BASNet/issues/38#issuecomment-630771144>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORKIZTHLV7NAPOQ5OSTRSJX3FANCNFSM4NDL3GVA .

OK, thank you for your patient and detailed answers. Thank you from the bottom of my heart.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/NathanUA/BASNet/issues/38#issuecomment-632418094, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORMZ7EJRH7AGBUBNP5LRSXDS7ANCNFSM4NDL3GVA .

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/