search

see-- / keras-centernet

A Keras implementation of CenterNet with pre-trained model (unofficial)
MIT License
337 stars 84 forks source link

DLA-34 model #7

Closed sonctmirasoft closed 5 years ago

sonctmirasoft commented 5 years ago

Hi, do you have any plan to implement the DLA-34 (deep layer aggregation) network described in the paper? I think it provide a better speed/accuracy trade off than the hourglass model (3 times faster for 3 to 4 less AP on COCO dataset).

see-- commented 5 years ago

Hi @sonctmirasoft ,

yes, I looked into DLA models. The problem is that TensorFlow does not support deformable convolutions. As you read, all other networks need them. There are open source libraries for deformable convolutions, but they aren't optimized. I don't think that you get 3x speedup. The hourglass model is probably still faster.

sonctmirasoft commented 5 years ago

Thanks for the head up, I guess we have to use the Pytorch version for now.

zyc4me commented 5 years ago

@see-- do you have the plan to release the training code？

bhack commented 4 years ago

You can upvote here for deformable convolution v2 https://github.com/tensorflow/addons/issues/179