search

blue-oil / blueoil

Bring Deep Learning to small devices
https://blueoil.org
Apache License 2.0
248 stars 86 forks source link

implement YOLOv3 based quantized object detection network #422

Open tkng opened 5 years ago

tkng commented 5 years ago

We're base network architecture for object detection is still YOLOv2. YOLOv3 looks much better than YOLOv2.

yd8534976 commented 5 years ago
Screen Shot 2019-09-13 at 15 42 53

We may have to give up coordinate refinement branch for this layer.

yd8534976 commented 5 years ago

If there is nobody implementing YOLOv3. I will take over this task but with tensorflow 2.0 (eager mode).

tkng commented 5 years ago

Thanks to riging your hand. Hoever, upgrading Blueoil's dependency of TF to 2.0 would be a complex project. Also, implementing YOLOv3 is also not an easy task. I don't recomment to do two complex things at the same time.

In this case, if you want to use TF 2.0, upgrading to TF 2.0 should be high priority.

tkng commented 5 years ago

Last but not least, before starting the upgrade to TF 2.0, we should consider how valuable it is. (Probably it would worth, but we don't know the detail yet.)

yd8534976 commented 5 years ago

@tkng Thanks for comments. I think a simple image classification task is not enough to evaluate the value of TF 2.0 for us. So I want to report does it worth by doing this task.. If you are available, let's have a discussion during CTO office hour.