EfficientDet?

[pwais]

Will an EfficientDet implementation be added to this repo? https://arxiv.org/pdf/1911.09070.pdf

From the paper, it sounds like it might be a straightforward addition to the detection module.

[pwais]

rando pytorch impl: https://github.com/SweetyTian/efficientdet/blob/master/bifpn.py

[pwais]

rando Keras impl https://github.com/xuannianz/EfficientDet

[2696120622]

Does EfficientDet require control flow like MASKRCNN?

[bhack]

An "anchor-free" version is available at https://github.com/xuannianz/SAPD

See also tensorflow/hub#424

[mingxingtan]

Official code is here: https://github.com/google/automl/tree/master/efficientdet

[pwais]

@mingxingtan Wish you guys were a lot more consistent with putting things in the same place and using the same harnesses, etc

[bhack]

@pwais It was already requested in the official Tensorflow Model tensorflow/models#6996

[mingxingtan]

We do wish to unify all models to use TF2.x in tensorflow/models, but unfortunately the
TF1 -> TF2 conversion is behind schedule, so we decide to first release TF1 version of EfficientDet for now.

[pwais]

@mingxingtan Ok sure, but then please re-open this issue? Or I'm happy to re-create a new one. It's nice that an alternative is available, and if you take 10 seconds to read this issue you'll see most of it is people linking to alternatives. But the ask of this issue is pretty straightforward: EfficientDet for this repo, and your post on twitter does not close it. I am, in particular, asking for you to be more considerate of those of us who don't work at Google.

[saberkun]

For the detection part, we actually need to check the owners:
@pengchongjin
If you would like to you the detection framework with EfficientDet

[mingxingtan]

@pwais Good point. Reopen it now.