Computer Vision News - November 2019

2 Summary Rese rch 6 The best property of AR models is that they provide a quantitative way to measure the model quality, i.e. Maximum Likelihood Estimator (MLE). The paper compares its performance on two unconditional generation benchmarks, CelebAHQ-256 and ImageNet of various sizes up to 256 . The authors demonstrate state of the art results (in terms of MLE score) on the ImageNet data set on various sizes. The table below shows the negative log likelihood score (smaller is better) of the suggested method compared to other AR methods: Another crucial aspect of the model is the order of slices prediction. In contrast to the conventional generation ordering that orders the pixels by the trivial order, the authors suggest an alternative ordering. This ordering divides a large image into a sequence of equally sized slices, where the pixel's location is also determined by taking into consideration the upscaling scheme. A scaling factor S is selected and each slice of size H/SxW/S is obtained by selecting a pixel every S pixels in both, height and depth. The suggested ordering is best illustrated in the following figure ((d),(e),(f),(g)).