Deformable neural radiance fields
Techniques of image synthesis using a neural radiance field (NeRF) includes generating a deformation model of movement experienced by a subject in a non-rigidly deforming scene. For example, when an image synthesis system uses NeRFs, the system takes as input multiple poses of subjects for training data. In contrast to conventional NeRFs, the technical solution first expresses the positions of the subjects from various perspectives in an observation frame. The technical solution then involves deriving a deformation model, i.e., a mapping between the observation frame and a canonical frame in which the subject's movements are taken into account. This mapping is accomplished using latent deformation codes for each pose that are determined using a multilayer perceptron (MLP). A NeRF is then derived from positions and casted ray directions in the canonical frame using another MLP. New poses for the subject may then be derived using the NeRF.
1 . A method, comprising:
acquiring image data representing a plurality of images, each of the plurality of images including an image of a scene within an observation frame, the scene including a non-rigidly deforming object viewed from a respective perspective;
generating a deformation model based on the image data, the deformation model describing movements made by the non-rigidly deforming object while the image data was generated, the deformation model being represented by a differentiable non-linear mapping between a position in the observation frame to a position in a canonical frame; and
generating a neural radiance field based on positions and viewing directions of casted rays through the positions in the canonical frame, the neural radiance field providing a mapping between the positions and viewing directions to a color and optical density at each position in the observation frame, the color and optical density at each position in the observation frame enabling a viewing of the non-rigidly deforming object from a new perspective.
2 . The method as in claim 1 , wherein the deformation model is conditioned on a latent code, the latent code encoding a state of the scene in a frame.
3 . The method as in claim 1 , wherein the deformation model includes a rotation, a pivot point corresponding to the rotation, and a translation.
4 . The method as in claim 3 , wherein the rotation is encoded as a pure log-quaternion.
5 . The method as in claim 3 , wherein the deformation model includes a sum of (i) a similarity transformation on a difference between a position and the pivot point, (ii) the pivot point, and (iii) the translation.
6 . The method as in claim 1 , wherein the deformation model includes a multilayer perceptron (MLP) within a neural network.
7 . The method as in claim 6 , wherein an elastic loss function component for the MLP is based on a norm of a matrix representing the deformation model.
8 . The method as in claim 7 , wherein the matrix is a Jacobian of the deformation model with respect to the position in the observation frame.
9 . The method as in claim 7 , wherein the elastic loss function component is based on a singular value decomposition of the matrix representing the deformation model.
10 . The method as in claim 9 , wherein the elastic loss function component is based on a logarithm of a singular value matrix resulting from the singular value decomposition.
11 . The method as in claim 7 , wherein the elastic loss function component is composed with a rational function to produce a robust elastic loss function.
12 . The method as in claim 6 , wherein a background loss function component involves designating points in the scene as static points that have a penalty for moving.
13 . The method as in claim 12 , wherein the background loss function component is based on a difference between a static point and a mapping of the static point in the observation frame to the canonical frame according to the deformation model.
14 . The method as in claim 6 , wherein generating the deformation model includes:
applying a positional encoding to a position coordinate within the scene to produce a periodic function of position, the periodic function having a frequency that increases with training iteration for the MLP.
15 . The method as in claim 14 , wherein the periodic function of the positional encoding is multiplied by a weight indicating whether a training iteration includes a particular frequency.
16 . A computer program product comprising a nontransitive storage medium, the computer program product including code that, when executed by processing circuitry of a computing device, causes the processing circuitry to perform a method, the method comprising:
acquiring image data representing a plurality of images, each of the plurality of images including an image of a scene within an observation frame, the scene including a non-rigidly deforming object viewed from a respective perspective;
generating a deformation model based on the image data, the deformation model describing movements made by the non-rigidly deforming object while the image data was generated, the deformation model being represented by a differentiable non-linear mapping between a position in the observation frame to a position in a canonical frame; and
generating a neural radiance field based on positions and viewing directions of casted rays through the positions in the canonical frame, the neural radiance field providing a mapping between the positions and viewing directions to a color and optical density at each position in the observation frame, the color and optical density at each position in the observation frame enabling a viewing of the non-rigidly deforming object from a new perspective.
17 . The computer program product as in claim 16 , wherein the deformation model includes a multilayer perceptron (MLP) within a neural network.
18 . The computer program product as in claim 17 , wherein an elastic loss function component for the MLP is based on a norm of a matrix representing the deformation model.
19 . The computer program product as in claim 18 , wherein the matrix is a Jacobian of the deformation model with respect to the position in the observation frame.
20 . An electronic apparatus, the electronic apparatus comprising:
memory; and
controlling circuitry coupled to the memory, the controlling circuitry being configured to:
acquire image data representing a plurality of images, each of the plurality of images including an image of a scene within an observation frame, the scene including a non-rigidly deforming object viewed from a respective perspective;
generate a deformation model based on the image data, the deformation model describing movements made by the non-rigidly deforming object while the image data was generated, the deformation model being represented by a differentiable non-linear mapping between a position in the observation frame to a position in a canonical frame; and
generate a neural radiance field based on positions and viewing directions of casted rays through the positions in the canonical frame, the neural radiance field providing a mapping between the positions and viewing directions to a color and optical density at each position in the observation frame, the color and optical density at each position in the observation frame enabling a viewing of the non-rigidly deforming object from a new perspective.