The result is not better than the post-processing CRF?

[zhangrui0828]

Thanks for your repo!!! I am applying your crfasrnn with deeplab-v2 for outdoor scene segmentation and compare it the post-processing CRF. Accoring to the introduction of the original auto (Suai Zhen, Sadeep J.), crfasrff is much better than the post-processing CRF, but I get the same results between crfasrnn and post-processing CRF. I am confused by this. For the CRF parameter set(theta_alpha, theta_beta, theta_gamma), are there any reqirements? Does crfasrnn have pretrained parameters?

Many thanks Rui

[Fettpet]

Accoring to the introduction of the original auto (Suai Zhen, Sadeep J.)

I dont know which source you mean.

but I get the same results between crfasrnn and post-processing CRF

What are the results without post processing? Also in this implementation is a bug. I dont know how this effects the performance. For me crfasrnn has improve the performance, two years ago . I havn't compare it with densecrf.

For the CRF parameter set(theta_alpha, theta_beta, theta_gamma), are there any reqirements?

They must be different from 0. Take a look at Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials

Does crfasrnn have pretrained parameters?

I think you can use the pretrained parameter of the neural network you use.

[zhangrui0828]

Thank you for your quick reply. Yes. I get the same results between crfasrnn and post-processing CRF. This is the problem!!! I think the result of crfasrnn should be much better than post-processing CRF. According to the paper(Conditional Random Fields as Recurrent Neural Networks), it should be improved about 3.4%. But the results of crfasrff I got was not better than that of post-processing CRF. I don't know why.

[netw0rkf10w]

I confirm that this is due to the bug reported in #6. Please check my code for a working implementation: https://github.com/netw0rkf10w/CRF.