Computer Vision News - November 2019

2 Summary Rese rch 8 There are also results on ImageNet data set generation. In general, the generation of ImageNet images is considered harder since the data distribution in this dataset is much more complicated. You can see the results below: The paper presents a new model for generating high fidelity images. It is applied with an incremental approach; first generating small images, then upscaling them in size, and finally, upscaling them in depth. To do so, the authors define a new architecture named Sub Pixel Network that provides a prediction for the next slice in the sequence given the previous slices. The paper demonstrated state of the art results in generating high fidelity images both in terms of MLE score and in qualitative results. Conclusion