goals.txt

SinGAN: Learning a Generative Model from a Single Natural Image
http://openaccess.thecvf.com/content_ICCV_2019/papers/Shaham_SinGAN_Learning_a_Generative_Model_From_a_Single_Natural_Image_ICCV_2019_paper.pdf

Qualitative: The most simple generative use, sampling random alternative images to the image used in training such as Figure 1 and Figure 6. 
If this is not acceptable as a minimum, we can reproduce the 'Editing' application as well, which is shown in Figure 12.

Quantitative: Table 2, we plan to reproduce the SIFID results (shown in the second column). We cannot show the correlation with the surveys since we will not be able to perform our own. 

---- version 1 submission ----
* We did not have to make any changes in our goals. We did not reproduce the 'Editing'
    application as it was an extra in case our main goals were deemed insufficient.

* We were able to train successful models on multiple images (the ones shown in the paper). We
    found that the random samples generated by our trained models were similar to the ones
    shown in Figures 1 and 6 in the paper. Our models can also generate samples with arbitrary
    output sizes, once again as illustrated in the paper.

    We also implemented SIFID as explained by the paper and validated our SIFID implementation
    over their survey dataset, whose results are shown in Table 2 in the paper. 
    However, as this dataset contains hand picked samples from fifty different trained 
    models, we did not fully reproduce their results. Instead, we computed the average 
    SIFID over 50 samples from each of the 8 models we trained, and found that their 
    average is very similar to the results reported in the paper.

    All of these results are illustrated in our notebook's saved outputs.

* We believe that we managed to reproduce the results we aimed to reproduce well enough :)

---- version 2 submission ----
* Since we had already achieved our goals, we only had to add the paper and code flow summary
    to our notebook and we did not make any changes in the code. Everything we've said under 
    version 1 submission still holds.