I don't think the RoiPoolConv layer is performing a max pool but rather an average pool

[trzy]

Great repo and well written blog! I have one question: RoiPoolConv uses tf.image.resize_images, which performs a bilinear interpolation, not a max of the pixels (unless I am misinterpreting its use here). I notice this was present in the original source repo (although that code also featured a Theano backend version that performed max-pooling correctly). This might be something to fix or at least call out with a comment.

[mehnaz1985]

@trzy, do you successfully run this repository? in the testing phase I am facing a problem, no boxes are created which means no objects are detected and also AP is 1 for all classes. could you pls respond to my query?

[trzy]

I haven’t tried. I’m writing my own FasterRCNN implementation from scratch using Keras that follows the paper more closely. I think my RPN is working but until I add the classifier layer sometime this week or next, I won’t know for sure. Feel free to reach out in a few weeks and see where I am at.

Sent from my iPhone

On Apr 13, 2021, at 6:29 PM, mehnaz1985 @.***> wrote:

 @trzy, do you successfully run this repository? in the testing phase I am facing a problem, no boxes are created which means no objects are detected and also AP is 1 for all classes. could you pls respond to my query?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[mehnaz1985]

Thanks for the reply.

Mehnaz Tabassum PhD Candidate School of SEIT Federation University, Gippsland Campus Victoria, Australia.

On Wed, Apr 14, 2021 at 12:33 PM trzy @.***> wrote:

I haven’t tried. I’m writing my own FasterRCNN implementation from scratch using Keras that follows the paper more closely. I think my RPN is working but until I add the classifier layer sometime this week or next, I won’t know for sure. Feel free to reach out in a few weeks and see where I am at.

Sent from my iPhone

On Apr 13, 2021, at 6:29 PM, mehnaz1985 @.***> wrote:

 @trzy, do you successfully run this repository? in the testing phase I am facing a problem, no boxes are created which means no objects are detected and also AP is 1 for all classes. could you pls respond to my query?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/RockyXu66/Faster_RCNN_for_Open_Images_Dataset_Keras/issues/79#issuecomment-819176997, or unsubscribe https://github.com/notifications/unsubscribe-auth/AP2WSD7BQBPH23DVLTJKRHLTIT5F3ANCNFSM4YLLOTDQ .

[giorgosrigas]

Thanks for the reply. Mehnaz Tabassum PhD Candidate School of SEIT Federation University, Gippsland Campus Victoria, Australia. … On Wed, Apr 14, 2021 at 12:33 PM trzy @.> wrote: I haven’t tried. I’m writing my own FasterRCNN implementation from scratch using Keras that follows the paper more closely. I think my RPN is working but until I add the classifier layer sometime this week or next, I won’t know for sure. Feel free to reach out in a few weeks and see where I am at. Sent from my iPhone > On Apr 13, 2021, at 6:29 PM, mehnaz1985 @.> wrote: > >  > @trzy, do you successfully run this repository? in the testing phase I am facing a problem, no boxes are created which means no objects are detected and also AP is 1 for all classes. could you pls respond to my query? > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub, or unsubscribe. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#79 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AP2WSD7BQBPH23DVLTJKRHLTIT5F3ANCNFSM4YLLOTDQ .

Hello. Did you figure your issue out? I have the same problem, no boxes drawn on the testing phase and three of my four classes have AP = 1.0 on each image. I am really lost here.