CVPR Daily - Wednesday

DAILY Wednesday ScanNet Indoor Scene Understanding 19 Andreas Geiger, professor at the University of Tübingen, will tell us about “Occupancy Networks”, which use neural networks to represent implicit surfaces by predicting whether a point or voxel is on the inside or outside of a shape. In his work at this year’s CVPR, Convolutional Occupancy Networks, he introduces convolutions to “Occupancy Networks” to help capture local information and structure for improved representation of 3D scenes. He will also describe how to represent texture and light information so that reconstructed 3D models can be accurately textured and lighted. By attending his talk you will also find out more about how we can create better reconstructions of real-world objects and scenes by performing joint estimation of pose, geometry, and material models. Yasutaka Furukawa , associate professor at Simon Fraser University and also recipient of the 2007 PAMI Longuet-Higgins Prize (congratulations Yasu!), will give a talk onwhy he believes “ CVPR is like a contemporary art exhibit ” and how great ideas can propagate through papers even if initially a particular paper does not have a large impact. Hewill also talk about howconvolutions can capture spatial information in graph neural networks for reconstructing outdoor architecture and generating floor plans. Thomas Funkhouser , senior staff research scientist at Google and professor emeritus at Princeton University, will tell us about recent work on semantic scene reconstruction. Besides these exciting speakers, we will also feature talks by winners of the ScanNet benchmark challenge. The ScanNet benchmark challenge consists of 3D semantic labeling and 3D instance segmentation tasks, where an input 3D scan (specified by vertices i.e. points) is labeled with semantic category or instances (see Figure from the ScanNet benchmark, with colors indicating semantic classes for each object, and the boxes enclose each object instance). This year, the workshop welcomes two speakers ( Chris Choy , research scientist at Nvidia and Benjamin Graham , research scientist at Facebook AI Research), whose sparse convolutional