ORN is very slow and backward is buggy.

[samson-wang]

Should do some formula reasoning first.

[MiZhangWhuer]

Hello, I also re-implement ORN caffe version. But I am not sure how to tackle with ORNAlign layer. Would you mind to further communicate with me?

[samson-wang]

@MiZhangWhuer I'd love to. However, I have only implemented ORConv and ORPooling. My version didn't work well on cifar, ether accuracy or performance. And I didn't go any further.

The good news is that Yanzhao has released his work. The most efficient way to reimplement ORN on caffe is to read his code. Or you can wait for the official caffe support as the README said.

Yanzhao has not released the result on Imagenet. I think there is a possibility that the orconv can't work well on large image. Although the results on cifar and mnist are impressive.

[MiZhangWhuer]

@samson-wang I have read the original source code based on Torch. But what confused me most is ROAlign part in their code. After finding the dominate orientation, they rotate each channels (orientation) by the index instead of coordinate interpolation. I have implemented ROAlign layer in the similar way as ArgmaxLayer and added back propagate process. But I am still not sure about whether the implementation is correct. How can I contact you online?

[samson-wang]

@MiZhangWhuer samson.c.wang@gmail.com

[samson-wang]

Tips:
From the official torch code, to speed up the forward, a customer im2col should be implemented. group convolution is really slow.