search

yun-liu / RCF

Richer Convolutional Features for Edge Detection
Other
758 stars 260 forks source link

BSDS Alone Performance #106

Open balajiselvaraj1601 opened 4 years ago

balajiselvaraj1601 commented 4 years ago

We took the code from https://github.com/meteorshowers/RCF-pytorch and were able to fix some performance issues and were able to reach the following results on BSDS500

ODS=0.794 OIS=0.811 AP=0.780 R50=0.859 - HED

We have trained the data only on BSDS dataset and analyzed only the single view results. By single view, i mean fused results from the last layer of the network without carrying out Multi scale testing.

Is this the best score possible ?

We are trying to replicate your results on BSDS dataset for our task

yun-liu commented 4 years ago

Sorry, I can't understand. Is your performance obtained using RCF or HED (I see HED in your results)? Does the last layer mean conv5_3? Why did you only use the last layer? Please refer to our paper for the results with only BSDS500 for training.

balajiselvaraj1601 commented 4 years ago

Thanks for the reply.  There were some typo errors. I will clearly explain the procedure we  carried out.

Our Results only using  BSDS500 dataset ODS=0.794 OIS=0.811 AP=0.780 R50=0.859

We took the code from https://github.com/meteorshowers/RCF-pytorch and were able to fix some performance issues.

By last layer, we meant the output from the fuse layer. It's normal approach used for getting single scale results as per your model.

Is this the best score possible with the BSDS alone training.

Repo for our code https://github.com/balajiselvaraj1601/RCF_Pytorch_Updated

yun-liu commented 4 years ago

Our Caffe code can achieve the performance of ODS 0.798 and IOS of 0.815 as shown in our paper, so I think your performance is very close to this.

balajiselvaraj1601 commented 4 years ago

Are you resizing your images to any specific dimension during training or testing ?

yun-liu commented 4 years ago

No

ForawardStar commented 3 years ago

@yun-liu ODS-F 0.798? The paper says that the performance of ODS is 0.806. I am confuse about it. Now I try to reproduce the results repoted in the paper, I do not know which one is the correct result.

yun-liu commented 3 years ago

@ForawardStar Sorry, I don't understand what your question is.

For the results, as shown in the paper, RCF trained with only the BSDS dataset achieves 0.798 ODS F-measure, and RCF trained with the BSDS + PASCAL datasets achieves 0.806 ODS F-measure (the standard setting).

ForawardStar commented 3 years ago

@yun-liu I got it. Thanks for your answer. One more question,if I want to train the model with BSDS+PASCAL, should I download the dataset from http://mftp.mmcheng.net/liuyun/rcf/data or somewhere else?

yun-liu commented 3 years ago

@ForawardStar Yes