CVPR Daily - Friday

The large-scale ImageNet database contains a million images with 1,000- way classification labels. These labels will show if an image features a dog or a cat, for example. Pixel-wise labeling, which is used in downstream tasks such as semantic segmentation and instance segmentation to know an object’s shape or to distinguish one thing from another in an image, is not present. Pixel-wise labeling is expensive and hard to obtain at scale. This work proposes leveraging a GAN with minimal manual labeling effort to synthesize a data set with images and pixel-wise labeling . “ If you’re working on a very practical application such as autonomous driving, you want to collect lots of images and need significant human effort for labeling, ” Daiqing explains. “ Usually, images are labeled pixel-wise, but even the largest data set is still very small. It’s just not practical enough to apply to a real real-world scenario. In our method, we can train a GAN on 1 million unlabeled images and Daiqing Li is a Research Scientist at the NVIDIA Toronto AI Lab. His paper proposes repurposing ImageNet GANs as data set generators. He speaks to us ahead of his poster presentation today. BigDatasetGAN: Synthesizing ImageNet with Pixel-Wise Annotations 14 DAILY CVPR Friday Poster Presentation