ECCV 2020 Daily - Monday

3 CoReNet: Stefan Popov 9 DAILY M o n d a y function 3D volume representation , like occupancy networks, and combined them with voxel grids to form a hybrid 3D volume representation. This enables building translation equivariant 3D models using standard convolutional blocks, while at the same time encoding fine object details without the excessive memory footprint usually associated with models based on voxel grid representations. It predicts the occupancy for a single point inside each voxel and then does this for the whole voxel grid. That way it can reconstruct a model with a much higher resolution than it has been trained for. object, the team proposed three technical extensions. You can learn more about these by viewing their presentation video. The first element was ray-traced skip connections that connect pixels in the input image to their corresponding frustum in the 3D volume. They allow the model to propagate occlusion boundaries and object contact points detected on the 2D image into 3D, and to understand the depth relations among objects locally. For the second element, the team took inspiration from works on implicit