CVPR Daily - Friday

and then ask a human annotator to label a few examples sampled from this generative model. Then we train a feature interpreter to interpret the labels from the latent space. As this is a generative model, we can do the sampling, and those labels will propagate to all the other classes or new samples. We can synthesize large-scale data sets using a generative model along with this interpreter. ” The work proposes only background/foreground segmentation for the 1,000 classes, but the team is already working on adding detailed parts, like labeling for eyes, noses, and animal features, which will help enable many downstream applications. Daiqing points out that relying on generative modeling is also a limitation of this work. Rare classes in ImageNet cannot be modeled by the GAN so well, so this method will not be able to synthesize them sufficiently. However, with the field moving so fast, he is sure GANs will soon get stronger and be able to model better data sets. The NVIDIA Toronto AI Lab team , where Daiqing is based, works in many exciting research directions. Recently, he has been focusing on leveraging generative models for discriminative tasks. Also, on top of their work on generative modeling and diffusion models, they are developing a neural rendering simulation for autonomous vehicles to add realistic hazards to make the simulation more useful. We ask Daiqing what it is like to work with co-author Sanja Fidler , Vice President of AI Research at NVIDIA and an Associate Professor at the University of Toronto. 15 DAILY CVPR Friday Daiqing Li