Maybe the network didn't learn the horizontal information

[Mr-doooog]

I trained a mobilenet v2 with the generated rotated image, which show great performance on generated rotated image but low performance on real rotated image. I think maybe the net learned the way to generate rotated images by interpolating artifacts or other methods

[vitorvilela]

I also faced that problem and thought the same as you. After I trained on "manually" rotated images I could get a model that worked as expected.

[KazuhideMimura]

Hi, I've made an extension to generate rotated images in smaller number of classes. If you don't need to determine the exact angle, please try a 4-step rotation (in 90° increments), which doesn't require to fill in the outside of the original image.

https://github.com/d4nst/RotNet/issues/50

[willswordh]

@vitorvilela Hi, could you share your model weight h5 file? Thanks a lot!

[vitorvilela]

Hi @willswordh,

I'm quite sure the model I trained will not work for you, because it was trained on a very specific kind of image. Unfortunately, I can't share the binary because of copyright.

Vitor Vilela

On Tue, Nov 21, 2023 at 4:36 PM willswordh @.***> wrote:

@vitorvilela https://github.com/vitorvilela Hi, could you share your model weight h5 file? Thanks a lot!

— Reply to this email directly, view it on GitHub https://github.com/d4nst/RotNet/issues/40#issuecomment-1821564530, or unsubscribe https://github.com/notifications/unsubscribe-auth/AC67HCA457YAPUSZMSHVCLTYFT7FHAVCNFSM5HK6VWJKU5DIOJSWCZC7NNSXTN2JONZXKZKDN5WW2ZLOOQ5TCOBSGE2TMNBVGMYA . You are receiving this because you were mentioned.Message ID: @.***>