Classification-of-object-in-a-specific-image-region-using-a-convolutional-neural-net-with-ROI-mask-a
Classification-of-object-in-a-specific-image-region-using-a-convolutional-neural-net-with-ROI-mask-a copied to clipboard
sagieppel
Baseline accuracy with no mask?
Hi Sagi,
I came across your paper https://arxiv.org/pdf/1812.00291.pdf as I have been doing some research on how to best combine a masked image with non-masked image. The findings in your report look interesting, but there is no comparison to the baseline accuracy when no mask is used. Do you happen to have that information?
Yes, merging the attention map with the first layer seem to perform much better than only using hard attention (zeroing out all regions outside the ROI in the image), but how does the former compare to the accuracy of only classifying the image as it is without using any masking?
Only checked this on cases where there are many different objects in the image, in that case, it will be impossible for a simple classification net to know which object to classify without the attention mask as input.
On Mon, Aug 29, 2022 at 7:50 AM Muhammad Alsharif @.***> wrote:
Hi Sagi,
I came across your paper https://arxiv.org/pdf/1812.00291.pdf as I have been doing some research on how to best combine a masked image with non-masked image. The findings in your report look interesting, but there is no comparison to the baseline accuracy when no mask is used. Do you happen to have that data?
Yes, merging the attention map with the first layer seem to perform much better than hard attention (zeroing out all regions outside the ROI in the image), but how does the former compare to the accuracy of only classifying the image as it is without using any masking?
— Reply to this email directly, view it on GitHub https://github.com/sagieppel/Classification-of-object-in-a-specific-image-region-using-a-convolutional-neural-net-with-ROI-mask-a/issues/4, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB37V535QPWTI3K5YHJ7VFLV3Q6R5ANCNFSM574LNBPA . You are receiving this because you are subscribed to this thread.Message ID: <sagieppel/Classification-of-object-in-a-specific-image-region-using-a-convolutional-neural-net-with-ROI-mask-a/issues/4 @github.com>
