huizeng / grid-anchor-based-image-cropping Goto Github PK

Hello Hui, in every txt file in the dataset you provided, there are always some crops that have scores of -2.00. Is that a bug in the dataset collection procedure?

when I run the demo_eval.py on ubuntu,I met an error:
ImportError:My_path/pytorch_cropping_evaluation/roi_align/_ext/roi_align/_roi_align.so: undefined symbol: PyInt_FromLong

how can I solve the problem
I use python3.6 pytorch0.4.1

Hello , you've done a good job! But there are no training details in pytorch code. So could you mind provide it? thanks a lot!

For generating crops having fixed aspect ratio, you need to change the generate_crop function to generate a set of candidate crops having fixed aspect ratio. You can refer to the implementations in testGAIC_qualitative_customer.m.

Originally posted by @HuiZeng in #10 (comment)

Hi Huizeng, I recently looked at the latest GAIC dataset (Grid Anchor Based Image Cropping: A New Benchmark and An Efficient Model).
I am confused by the fact that this dataset has greater than 0 and less than 1 in the score labels, e.g. 0.86. Also, regarding this latest dataset, I also checked the distribution of scores and I got the result in the following figure, which is different from Figure7 that you gave in your paper. Looking forward to your reply.

I not find the url to download the dataset GAICD( Grid Anchor
based Image Cropping Database)，is there a way to down the dataset？

I don't run you pytorch project becouse i have an error

ImportError: torch.utils.ffi is deprecated. Please use cpp extensions instead.

Before I run this code,I setup and test the matconvent successfully.But I run the testGAIC_qualitative_customer.m for evulation,I saw these error information.

`Undefined function or variable 'dagnn.RoiAlign'.

Error in dagnn.DagNN.loadobj (line 26)
block = constr() ;

Error in testGAIC_qualitative_customer (line 19)
net = dagnn.DagNN.loadobj(netStruct.net) ;

Error in run (line 91)
evalin('caller', strcat(script, ';'));
`

I trying to implement your paper in tensorflow and run experiments in less models, but i get overfitting in training set. You encountered this problem in your experiment (I did not find this in the paper)?

Hello. I downloaded GAICD and saw annotations.
MOS is sometimes "-2" but rating is 1 to 5.
What does "-2" mean?

Such as 493063.txt
24 43 408 811 3.71
24 43 408 896 -2.00
24 43 408 981 -2.00
24 43 456 725 3.71
24 43 456 811 4.00

place(24,43) (408,896)
MOS -2.00

Is this "2"?
or lowest "1"?
or worst is "-2" actually?

Thank you for your answer

#18

Hi, I am running the pytorch version on my local environment(windows 10+pytorch 0.4+python3.6),got following error message:
ModuleNotFoundError: No module named 'roi_align._ext.roi_align._roi_align'

I found there is an .so file that windows can't read.
How can I solve this error?
Looking forward to hearing from you soon.

Hi , when I run the demo_eval.py on linux, there had an error:
ImportError :/pytorch_cropping_evaluation/roi_align/_ext/roi_align/_roi_align.so: undefined symbol: PyInt_FromLong

could you please tell me how can I solve the problem?

We recently found an implementation bug in calcualating the Acc indexes which were slightly different from the definition in our paper.
Fortunately, this bug does not affect any of the conclusions in our paper.
We have corrected this bug in the Matlab code and updated the corresponding values in the revised paper using blue-color font.
We apologize for the confusion to readers. The detailed changes in code can be found here:
5818679#diff-d2f9e0c4e433d155b38659652521d46a

The original code actually calculates the following Acc:

rather than the real definition in our paper:

Hello!

Thank you very much for sharing the code and dataset with the community. I have a questions about how would one go about testing previously published methods on the GAICD. I was wondering if you could possibly describe how do you test the VFN / VEN models on the GAICD dataset. Do you feed take the precomputed crops and then feed them each into the VFN / VEN and then rank order the crop scores?

Thank you very much.

I installed the python version code of this project. In README.md file, there is "Notification: the trained model naturally supports generating crops having fixed aspect ratios or resolutions. This function can be supported when necessary in the future.".
So how can I specific the ratio and resolutions of the output?

I want to increase the epoches of the train or the times of the train in order to improve the training model of my dataset.Or are there any other ways to improve the training model.

Maybe i meet the error.But I don't know what cause that.I run the testmodel.m and print the highest score of each photo.In the first 1-5 epoch,the highest score normally is 3-5.But in the later epoch,the highest score just 1.
I use the same dataset about the train and the test.But the testmodel.m can't give me the five point in the later epoch.It confused me.
Thank u for u reply.

Can i get your loss and acc curves during training?