Meta Platforms, Inc.

Meta Platforms Technologies, LLC

Jia-Bin Huang

Johannes Peter Kopf

Changil Kim

Michael Zollhoefer

Wenqi Xian

Xuejian Rong

Benjamin Attal

In particular embodiments, a computing system may access a set of training images for training a neural light field network for view synthesis. Using each training image, the computing system may train the neural light field network by casting, for each pixel of the training image, a ray into a three-dimensional (3D) space, the ray including integrated radiance along the ray, mapping first ray coordinates of the ray into an embedding network, transforming, using the embedding network, the first ray coordinates into second ray coordinates, applying positional encoding to second ray coordinates, generating, using the neural light field network, a predicted color value for the pixel based on positionally encoded second ray coordinates, comparing the predicted color value for the pixel with a ground-truth color value for the pixel, and updating the neural light field network and the embedding network based on comparison.

Learning neural light fields with ray-space embedding networks

In particular embodiments, a computing system may receive a video comprising a plurality of image frames. The system may generate, for each image frame in the video, an initial depth map using a machine-learning model. The system may compute a misalignment error indicating depth misalignments in initial depth maps using a reprojection technique. The system may generate, for each image frame in the video, an optimized camera pose and a flexible deformation spline associated with the image frame to minimize the misalignment error. The system may generate, for each image frame in the video, a refined depth map by adjusting the initial depth map associated with the frame using the flexible deformation spline associated with the image frame.

Robust consistent video depth estimation

In one embodiment, a computing system may determine a view position, a view direction, and a time with respect to a scene. The system may access a spatiotemporal representation of the scene generated based on (1) a monocular video including images each capturing at least a portion of the scene at a corresponding time and (2) depth values of the portion of the scene captured by each image. The system may generate an image based on the view position, the view direction, the time, and the spatiotemporal representation. A pixel value of the image corresponding to the view position may be determined based on volume densities and color values at sampling locations along the view direction and at the time in the spatiotemporal representation. The system may output the image to the display, representing the scene at the time as viewed from the view position and in the view direction.

Space-time representation of dynamic scenes

Growing community of inventors

Blacksburg, VA, United States of America

Jia-Bin Huang