search

tensorflow / models

Models and examples built with TensorFlow
Other
76.94k stars 45.79k forks source link

Add rotated bounding boxes and anchors for models #10893

Open Petros626 opened 1 year ago

Petros626 commented 1 year ago

Hello dear TensorFlow-team,

I wanted to ask if there are any plans to add a long awaited feature to the Object Detection API in the near future. It is about rotated bounding boxes for training and execution of models.

Nvidia has already taken the step here and the code for implementing this feature is certainly very helpful. (https://github.com/NVIDIA/retinanet-examples/blob/main/odtk/box.py)

On behalf of the community, we would all certainly be very happy to receive this feature.

Greetings,

Petros

laxmareddyp commented 1 year ago

@fyangf,

Could you please look into this issue in your free time.

Thanks.

fyangf commented 1 year ago

Object Detection API is now maintained by the community so you are welcome to contribute to this feature. If you are talking about the models under official, then we do not have a plan to support rotated box in the short term. However, if you really care about rotated box, you may take a look at centernet: https://github.com/tensorflow/models/tree/master/official/projects/centernet, which is anchor free and can naturally handle rotated boxes (some code change may be needed).

Petros626 commented 1 year ago

Object Detection API is now maintained by the community so you are welcome to contribute to this feature. If you are talking about the models under official, then we do not have a plan to support rotated box in the short term. However, if you really care about rotated box, you may take a look at centernet: https://github.com/tensorflow/models/tree/master/official/projects/centernet, which is anchor free and can naturally handle rotated boxes (some code change may be needed).

Hey, thanks for the reply.

In principle, I think it would make sense for Tensorflow to equip all models with this feature, but this requires renewed training (Coco dataset) of all models with the rotated boxes, which could be made available to the community. I lack the hardware and time to train for all common models from scratch.

So this feature will probably not appear, right?