Differentiable real-time radiance field rendering for large scale view synthesis
A method that includes obtaining images of an environment that are captured by one or more image capture devices, determining intrinsic parameters and extrinsic parameters of the one or more image capture devices that are associated with each of the images, creating a differentiable radiance field associated with the environment, and generating a three-dimensional representation of the environment. The three-dimensional representation contains one or more portions of the environment uncaptured in the images.
1 . A method comprising:
obtaining images of an environment that are captured by one or more image capture devices, wherein at least one of intrinsic parameters or extrinsic parameters is different between the two or more image capture devices;
determining the intrinsic parameters and the extrinsic parameters of the one or more image capture devices that are associated with each of the images, wherein the extrinsic parameters include spatial location coordinates and a viewing direction with respect to a reference image;
determining neural graphics primitives of the environment based upon the images of the environment, wherein the neural graphics primitives comprise learned representations containing information about geometric and appearance features of the environment extracted from the images using a neural network trained to optimize an underlying continuous three-dimensional representation function using a sparse set of input views of the environment;
creating a differentiable radiance field associated with the environment based upon the intrinsic parameters, the extrinsic parameters, and the neural graphics primitives,
wherein the differentiable radiance field is a neural representation that models three-dimensional scene properties of the environment, including color and density, as continuous functions optimized via differentiable optimization training to enable rendering from novel viewpoints;
generating, based upon the differentiable radiance field, a three-dimensional representation of the environment, wherein the three-dimensional representation contains one or more portions of the environment not captured in the images; and
generating additional synthetic images from the differentiable radiance field to augment a training set of the neural network for enabling the neural graphics primitives to be trained on additional viewpoints and reducing reconstruction inaccuracies caused by unrepresented regions.
2 . The method of claim 1 , further comprising:
prior to determining the intrinsic parameters and the extrinsic parameters of the one or more image capture devices, filtering the images based on one or more filters to create a filtered image set, wherein the one or more filters includes at least one of the following filters: a file dimension filter, an image shape filter, a duplicate image filter, a grayscale filter, a color histogram filter, or a blur detection filter; and
filtering the images based on one or more filters using threshold values to create a filtered image set, wherein the filtering improves accuracy of intrinsic and extrinsic parameter determination and quality of the three-dimensional representation.
3 . The method of claim 1 , wherein the one or more image capture devices comprises a first image capture device and a second image capture device, further comprising:
determining a shared reference image between at least one image captured by the first image capture device and at least one image captured by the second image capture device;
based upon the determined reference image, determining a relative position of the first image capture device and the second image capture device with respect to the shared reference image; and
combining the images from the first and second image capture devices to create a shared image set using the determined relative position; wherein the differentiable radiance field is created based on the shared image set.
4 . The method of claim 1 , wherein the extrinsic parameters include spatial location coordinates and a viewing direction of the one or more image capture devices associated with each of the images.
5 . The method of claim 1 , wherein the images of the environment captured by the one or more image capture devices are two-dimensional images.
6 . The method of claim 5 , wherein the images of the environment are captured by two or more image capture devices, and at least one of the intrinsic parameters or the extrinsic parameters is different between the two or more image capture devices.
7 . The method of claim 1 , wherein the intrinsic parameters and the extrinsic parameters are initially represented by a CAHVOR model or a pinhole camera model for each of the images.
8 . The method of claim 7 , wherein the intrinsic parameters and the extrinsic parameters are initially represented by the CAHVOR model, and wherein the method further comprises:
prior to creating the differentiable radiance field, converting the representation of the intrinsic parameters and the extrinsic parameters from the CAHVOR model to the pinhole camera model.
9 . The method of claim 1 , further comprising:
determining a quality of the three-dimensional representation compared to the environment; and
based on the quality determined, adjusting one or more hyper-parameters of the neural network that creates the differentiable radiance field.
10 . The method of claim 9 , wherein the quality is determined by determining a peak signal-to-noise ratio between the environment and the three-dimensional representation.
11 . The method of claim 1 , further comprising:
determining an uncertainty of the three-dimensional representation with respect to the one or more portions of the environment not captured in the images; and
generating, based upon the determined uncertainty, an uncertainty map.
12 . The method of claim 11 , wherein the uncertainty map is a visual representation of geometric and texture uncertainties with respect to the one or more portions of the environment not captured or under represented in the images.
13 . The method of claim 11 , wherein the uncertainty map is generated based on a statistical method and the differentiable radiance field is an optimized neural radiance field (NeRF) or a 3D Gaussian splatting.
14 . A device, comprising:
a non-transitory memory; and
a processor configured to execute instructions stored in the non-transitory memory to:
obtain images of an environment that are captured by one or more image capture devices, wherein at least one of intrinsic parameters or extrinsic parameters is different between the two or more image capture devices;
determine the intrinsic parameters and the extrinsic parameters of the one or more image capture devices that are associated with each of the images, wherein the extrinsic parameters include spatial location coordinates and a viewing direction with respect to a reference image;
determine neural graphics primitives of the environment based upon the images of the environment, wherein the neural graphics primitives comprise learned representations containing information about geometric and appearance features of the environment extracted from the images using a neural network trained to optimize an underlying continuous three-dimensional representation function using a sparse set of input views of the environment;
create, a differentiable radiance field associated with the environment based upon the intrinsic parameters, the extrinsic parameters, and the neural graphics primitives,
wherein the differentiable radiance field is a neural representation that models three-dimensional scene properties of the environment, including color and density, as continuous functions optimized via differentiable optimization training to enable rendering from novel viewpoints; and
generate, based upon the differentiable radiance field, a three-dimensional representation of the environment, wherein the three-dimensional representation contains one or more portions of the environment not captured in the images.
15 . The device of claim 14 , wherein the images of the environment are captured by a first image capture device and a second image capture device, and at least one of the intrinsic parameters or the extrinsic parameters is different between the first image capture device and the second image capture device.
16 . The device of claim 15 , wherein the images include:
a first image set that contains images of the environment captured by the first image capture device, and a second image set that contains images of the environment captured by the second image capture device, and
wherein the instructions further include instructions to:
determine a commonality between at least one of the images of the first image set and at least one of the images of the second image set; and
based upon the commonality determined, determine a relative position of the first image capture device with respect to the second image capture device for each of the images of the first image set and the second image set.
17 . The device of claim 16 , wherein after determining the relative position of the first image capture device with respect to the second image capture device for each of the images of the first image set and the second image set, the instructions further include instructions to:
combine the first image set together with the second image set to create a shared image set using the relative position determined, wherein the three-dimensional representation of the environment is created based upon the shared image set.
18 . A non-transitory computer-readable medium that stores computer-executable instructions that, when executed by a processor, perform operations comprising:
obtaining images of an environment that are captured by one or more image capture devices, wherein at least one of intrinsic parameters or extrinsic parameters is different between the two or more image capture devices;
determining the intrinsic parameters and the extrinsic parameters of the one or more image capture devices that are associated with each of the images, wherein the extrinsic parameters include spatial location coordinates and a viewing direction with respect to a reference image;
determining neural graphics primitives of the environment based upon the images of the environment, wherein the neural graphics primitives comprise learned representations containing information about geometric and appearance features of the environment extracted from the images using a neural network trained to optimize an underlying continuous three-dimensional representation function using a sparse set of input views of the environment;
creating a differentiable radiance field associated with the environment based upon the intrinsic parameters, the extrinsic parameters, and the neural graphics primitives,
wherein the differentiable radiance field is a neural representation that models three-dimensional scene properties of the environment, including color and density, as continuous functions optimized via differentiable optimization training to enable rendering from novel viewpoints; and
generating, based upon the differentiable radiance field, a three-dimensional representation of the environment, wherein the three-dimensional representation contains one or more portions of the environment not captured in the images.
19 . The non-transitory computer-readable medium of claim 18 , wherein the device is one of a personal device or a cloud computer, and the computer-executable instructions are included in a downloadable software application.
20 . The non-transitory computer-readable medium of claim 19 , wherein the downloadable software application is executable to implement at least one of a simulation procedure, a learning procedure, a machine learning procedure, or a differentiable procedure.