Abstract: Provided are machine learning models that generate geometry-free neural scene representations through efficient object-centric novel-view synthesis. In particular, one example aspect of the present disclosure provides a novel framework in which an encoder model (e.g., an encoder transformer network) processes one or more RGB images (with or without pose) to produce a fully latent scene representation that can be passed to a decoder model (e.g., a decoder transformer network). Given one or more target poses, the decoder model can synthesize images in a single forward pass. In some example implementations, because transformers are used rather than convolutional or MLP networks, the encoder can learn an attention model that extracts enough 3D information about a scene from a small set of images to render novel views with correct projections, parallax, occlusions, and even semantics, without explicit geometry.

Abstract: Systems and methods for view synthesis and three-dimensional reconstruction can learn an environment by utilizing a plurality of images of an environment and depth data. The use of depth data can be helpful when the quantity of images and different angles may be limited. For example, large outdoor environments can be difficult to learn due to the size, the varying image exposures, and the limited variance in view direction changes. The systems and methods can leverage a plurality of panoramic images and corresponding lidar data to accurately learn a large outdoor environment to then generate view synthesis outputs and three-dimensional reconstruction outputs. Training may include the use of an exposure correction network to address lighting exposure differences between training images.