search

fizyr / keras-retinanet

Keras implementation of RetinaNet object detection.
Apache License 2.0
4.38k stars 1.96k forks source link

Pyramid level customization #1381

Closed mooratov closed 4 years ago

mooratov commented 4 years ago

made pyramid levels configurable, and implemented the capability to use P2 pyramid layer. Currently, P2 layers only works for Resnet backbone.

mooratov commented 4 years ago

very sorry, this was supposed to go on my personal fork (not ready for primetime)

muaali commented 3 years ago

@mooratov I am trying to understand what is the benefit of using P2 and C2 in retinanet. Also did it benefit you in some way, in terms of performance. What was the scale / size of objects that you were trying to detect. Thanks.

mooratov commented 3 years ago

we were interested in very small bounding boxes, on the scale of 4x4 pixels or so.

On Tue, Mar 23, 2021 at 6:29 AM muaali @.***> wrote:

@mooratov https://github.com/mooratov I am trying to understand what is the benefit of using P2 and C2 in retinanet. Also did it benefit you in some way, in terms of performance. What was the scale / size of objects that you were trying to detect. Thanks.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/fizyr/keras-retinanet/pull/1381#issuecomment-804903466, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABJKFAR3TTW6J4DSPVHRU3DTFCJT7ANCNFSM4NRF6ZWA .

muaali commented 3 years ago

Thanks @mooratov . I will use your commits to include C2-P2 in the model and then to train the model and see if it learns better.