About the PSNR of wild(55) dataset during testing

[Rachel-kk]

Hi，I tested your released model in wild(55) dataset, but it got psnr of 25.09, ssim of 0.889. I don't understand why the psnr I tested is not the same as your paper(the psnr of 25.26, the ssim of 0.890 in your paper).

Do you have any other processing of the dataset?

[Rachel-kk]

Also, it is mentioned in your paper: "Instead, suppose that there are N training datasets, where the i-th dataset is denoted by Si, we train an SIRR model for each of the dataset."

When ERRNet is used as backbone network, how do you train on real90 and unaligned datasets?

[CastellanLiu]

Thanks for your interest in our work.

1) Because we use the domain-level expert weights (i.e., we average the $w_i$ for all samples in that test set), and you should change the --avg parameter for testing on other datasets (see options/base_options.py). You can print the weights (the "attn_" here) and calculate the average value by yourself (it should be the same as the pre-defined ones in our commented options). Sorry for the inconvenience.

2) ERRNet is a backbone architecture, we train three models (with ERRNet architecture) by i) initializing the model with the ERRNet official weights, and ii) fine-tuning them with these datasets (e.g., real90, unaligned).

Besides, there is a mistake in our code, i.e., the number of images in Wild (see here). It should be 55, not 50. However, it only influences the final average metrics calculated by this code. Fortunately, the values in the paper are correct since we only recorded the metrics for each dataset and calculated the final average metrics in Excel.

[Rachel-kk]

Hi, thanks for your reply.

1. According to your method, I change the --avg parameter for testing on my local datasets. However, I got psnr on wild and solid dataset which is not the same as your paper. The result is as follows:

	Real20 (20)	Wild (55)	Postcard (199)	Solid (200)
your paper	22.80 / 0.790	25.26 / 0.890	23.08 / 0.874	25.26 / 0.904
our test	22.80 / 0.790	25.16 / 0.889	23.08 / 0.874	25.17 / 0.899

Can you provide your test dataset?

[CastellanLiu]

Hi, fortunately, we can get access to the SIR dataset without request in advance now. Thus I uploaded our test sets to the Baidu net disk (https://pan.baidu.com/s/10BOHUyZGYt0c-9YqP7pVDw?pwd=kka2). Note that we have converted the .jpg files to .png files. Please have a try again and see whether it is because of the difference between the datasets.

[Rachel-kk]

Hi，Thank you so much!

I retested on your test datasets, and I got the same performance as your paper. However, I'm still a bit curious why have you converted the .jpg files to .png files?

[CastellanLiu]

Actually, the evaluation dataset I provided was prepared by my colleague (please refer to https://github.com/liyucs/RAGNet). I just use these datasets directly.

[Rachel-kk]

ok, thanks

I am going to reproduce your paper...

[Rachel-kk]

Hi,

When training an SIRR model (with ERRNet architecture) for each of the dataset (e.g., Real90, Unaligned, Syn_CEIL), I ran this command.

python train_errnet.py \
  --name errnet_real90  \
  --hyper \
  -r \
  --icnn_path ./checkpoints/errnet_060_00463920.pt \
               ./checkpoints/errnet_060_00463920.pt \
               ./checkpoints/errnet_060_00463920.pt \
  --nModel 3

After training, I ran this command to test.

python test_errnet.py \
  --name errnet_real90  \
  --hyper \
  -r \
  --icnn_path ./checkpoints/errnet_real90/errnet_latest.pt \
  --nModel 3

Is the above correct?

Also， when training expert model for Syn_CEIL, did you crop the center region with size 224 x 224 for VOC2012?

[CastellanLiu]

Sorry for the late response.

Actually, you should train the domain experts via the official code of ERRNet. I mean, train each of the domain experts separately. Then they are used as domain experts and trained with my code modified from ERRNet.

For cropping, I follow the official code of ERRNet (see https://github.com/csmliu/AdaNEC/blob/master/data/reflect_dataset.py#L62)

[Rachel-kk]

Hi, I ran three experiments by changing the training data. Unfortunately, none of the experiments reproduced the performance of your paper. The results are obtained using the testing dataset you provided, as follows:		Real(20)	Wild(55)	Postcard(199)	Solid(200)
paper	22.80/0.790	25.26/0.890	23.08/0.874	25.26/0.904
Exp_1	23.04/0.792	24.88/0.877	22.75/0.873	24.95/0.898
Exp_2	23.00/0.790	25.11/0.887	22.21/0.869	24.79/0.896
Exp_3	22.94/0.790	25.22/0.886	22.33/0.877	24.35/0.891

Exp_1:

1. Firstly, I trained the domain experts via the official code of ERRNet for three datasets separately.

   • Synthetic dataset: Before training, I used the center crop to 224x224 on the original VOC2012.
   • Real89 dataset: the Height of Real89 images is resized to 512 provided by the authors of the ERRNet. Then, it is cropped to 224x224 on the fly.
   • Unaligned250 dataset: It is cropped to 224x224 on the fly.

2. Then they are used as domain experts and trained with your code, the trained weights for the testing data are as follows:

   • Real20: [0.092550091, 0.063934185, 0.843515721]
   • wild: [0.083660906, 0.230148451, 0.686190650]
   • postcard: [0.001090158, 0.645903816, 0.353006029]
   • solid: [0.283300335, 0.388968509, 0.327731159]

Exp_2

1. Firstly, I trained the domain experts via the official code of ERRNet for three datasets separately.

   • Synthetic dataset: Before training, I cropped to 224x224 on the original VOV2012 by reflect_dataset.py.
   • Real89 dataset: the same as Exp_1
   • Unaligned250 dataset: the same as Exp_1

2. Then they are used as domain experts and trained with your code, the trained weights for the testing data are as follows:

   • Real20: [0.148800396, 0.044181681, 0.807017929]
   • wild: [0.223979196, 0.217365322, 0.558655489]
   • postcard: [0.061034970, 0.300128878, 0.638836153]
   • solid: [0.558580633, 0.181719502, 0.259699868]

Exp_3

1. Firstly, I trained the domain experts via the official code of ERRNet for three datasets separately.

   • Synthetic dataset: The same as Exp_2
   • Real89 dataset: The shorter side of Real89 images is resized to 512
   • Unaligned250 dataset: The same as Exp_1

2. Then they are used as domain experts and trained with your code, the trained weights for the testing data are as follows:

   • Real20: [0.079918947, 0.007921040, 0.912160021]
   • wild: [0.100883266, 0.125001486, 0.774115256]
   • postcard: [0.165105508, 0.211451088, 0.623443398]
   • solid: [0.433878908, 0.161532401, 0.404588691]

By the way, I trained the model on TITAN V Pytorch1.7.1+cuda10.1.

[CastellanLiu]

I would like to check several things below: 1) Did you initialize the model when training the domain experts? 2) If yes, which model did you use to initialize the model? 3) For ERRNet, the performance of the domain experts has been provided in Table 4. Before checking the final results, could you check your domain experts? 4) Are the domain expert parameters fixed during training? Actually, the training code in this repo is not well organized, I will release a better version later (but not in the recent days since I am busy recently).

[Rachel-kk]

1. Yes， I used the same model( ERRNet official weights) for initialization training all the domain experts. And each dataset is iterated 20 times.
2. No, I didn't check my domain experts before checking the final results. Now, I test my domain experts, and the model performance metrics for my three experiments are as follows.
3. I directly ran your training code using the domain experts without any modifications. I guess the domain expert parameters should be fixed.

[CastellanLiu]

Sorry, I used the model I finetuned with train_errnet_unalign.py for initialization. As mentioned in the readme, the training code is not exactly the one I used, it should work fine, but I'm not very sure about it.

[Rachel-kk]

Thank you for telling me these, And I will retrain the initial model in this way.

Looking forward to your new code!