Computer Vision News - February 2020

2 Summary Rese rch 6 Where C denotes the neural network and is the intermediate layer of VGG16 (taken from 3 different layers). RefineNet - this network aims to generate the final output of the model: practically, to refine the output of CoarseNet. It is trained using the same loss as the CoarseNet with an addition of an adversarial loss over its output. In this addition, an adversarial training scheme uses a conditional discriminator, the goal of which is to distinguish between real source images used to generate the SfM model and images synthesized by RefineNet. In summary, the RefineNet final loss is of the form: Where R is the RefineNet neural network and D is the discriminator we mentioned above. Φ = || ( ) − || 1 + ∑ ||Φ ( ( )) − Φ ( )|| 2 2 =1,2,3 + [log( ( )) + log (1 − ( ( )))] Results The paper contains a mass of quantitative and qualitative experiments. We will highlight a few of the results presented and further details can be found in the paper. The following figure demonstrates the quality of the final output of the model. Each 3 by 1 set of squares shows the point cloud, the reconstructed image, and the original image. Although this task is quite new, results are remarkable. In some of the reconstructed images, the fake images cannot be distinguished from the original image, meaning that the method worked very well.