Shanghaitech University

Xiamen University

Plex-vr Digital Technology (shanghai) Co., Ltd.

Jingyi Yu

Minye Wu

Yi Ma

Zhiru Shi

Lan Xu

Wenguang Ma

Ziran Xing

Jialian Li

Chenglu Wen

Huangjie Yu

Cheng Wang

Haimin Luo

Cen Wang

A deep neural network based hair rendering system is presented to model high frequency component of furry objects. Compared with existing approaches, the present method can generate photo-realistic rendering results. An acceleration method is applied in our framework, which can speed up training and rendering processes. In addition, a patch-based training scheme is introduced, which significantly increases the quality of outputs and preserves high frequency details.

Hair rendering system based on deep neural network

Described herein are systems and methods for training machine learning models to generate three-dimensional (3D) motions based on light detection and ranging (LiDAR) point clouds. In various embodiments, a computing system can encode a machine learning model representing an object in a scene. The computing system can train the machine learning model using a dataset comprising synchronous LiDAR point clouds captured by monocular LiDAR sensors and ground-truth three-dimensional motions obtained from IMU devices. The machine learning model can be configured to generate a three-dimensional motion of the object based on an input of a plurality of point cloud frames captured by a monocular LiDAR sensor.

System and method of capturing three-dimensional human motion capture with LiDAR

A method of processing light field images for separating a transmitted layer from a reflection layer. The method comprises capturing a plurality of views at a plurality of viewpoints with different polarization angles; obtaining an initial disparity estimation for a first view using SIFT-flow, and warping the first view to a reference view; optimizing an objective function comprising a transmitted layer and a secondary layer using an Augmented Lagrange Multiplier (ALM) with Alternating Direction Minimizing (ADM) strategy; updating the disparity estimation for the first view; repeating the steps of optimizing the objective function and updating the disparity estimation until the change in the objective function between two consecutive iterations is below a threshold; and separating the transmitted layer and the secondary layer using the disparity estimation for the first view.

Light field based reflection removal

An image-based method of modeling and rendering a three-dimensional model of an object is provided. The method comprises: obtaining a three-dimensional point cloud at each frame of a synchronized, multi-view video of an object, wherein the video comprises a plurality of frames; extracting a feature descriptor for each point in the point cloud for the plurality of frames without storing the feature descriptor for each frame; producing a two-dimensional feature map for a target camera; and using an anti-aliased convolutional neural network to decode the feature map into an image and a foreground mask.

Multi-view neural human rendering

According to some embodiments, an imaging processing method for extracting a plurality of planar surfaces from a depth map includes computing a depth change indication map (DCI) from a depth map in accordance with a smoothness threshold. The imaging processing method further includes recursively extracting a plurality of planar region from the depth map, wherein the size of each planar region is dynamically adjusted according to the DCI. The imaging processing method further includes clustering the extracted planar regions into a plurality of groups in accordance with a distance function; and growing each group to generate pixel-wise segmentation results and inlier points statistics simultaneously.

System and method for extracting planar surface from depth image

A method of rendering an object is provided. The method comprises: encoding a feature vector to each point in a point cloud for an object, wherein the feature vector comprises an alpha matte; projecting each point in the point cloud and the corresponding feature vector to a target view to compute a feature map; and using a neural rendering network to decode the feature map into a RGB image and the alpha matte and to update the feature vector.

Neural opacity point cloud

A method of detecting and recognizing faces using a light field camera array is provided. The method includes capturing multi-view color images using the light field camera array; obtaining a depth map; conducting light field rendering using a weight function comprising a depth component and a sematic component, where the weight function assigns a ray in the light field with a weight; and detecting and recognizing a face.

Face detection and recognition method using light field camera system

The present invention relates to an all-around spherical light field rendering method, comprising: a preparation step, i.e., preparing to input and load related files; pre-computing a latticed depth map of positions of reference cameras which are densely covered on a sphere; moving a rendering camera, enabling the moving range of the rendering camera to be the surface of the sphere, and calculating and identifying the reference cameras surrounding the rendering camera around the rendering camera; performing back projection on pixels of the rendering camera, and performing depth test with the four reference cameras; and interpolating the reference cameras passing through the depth test, thereby obtaining a finally rendered pixel value. By means of the present invention, the rendering results can be rapidly seen in real time; an object can be observed from any angle on the spherical surface, and a real immersion feeling can be felt.

All-around spherical light field rendering method

A method of generating a three-dimensional model of an object is disclosed. The method may use a light field camera to capture a plurality of light field images at a plurality of viewpoints. The method may include capturing a first light field image at a first viewpoint; capturing a second light field image at the second viewpoint; estimating a rotation and a translation of a light field from the first viewpoint to the second viewpoint; obtaining a disparity map from each of the plurality of light field image; and computing a three-dimensional point cloud by optimizing the rotation and translation of the light field and the disparity map. The first light field image may include a first plurality of subaperture images and the second light field image may include a second plurality of subaperture images.

Method and system for three-dimensional model reconstruction

A method of compressing a stereoscopic video including a left view frame and a right view frame is provided, the method including: determining a texture saliency value for a first block in the left view frame by intra prediction (); determining a motion saliency value for the first block by motion estimation (); determining a disparity saliency value between the first block and a corresponding second block in the right view frame (); determining a quantization parameter based on the disparity saliency value, the texture saliency value, and the motion saliency value (); and performing quantization of the first block in accordance with the quantization parameter ().

Compression method and apparatus for panoramic stereo video system

A method and apparatus for extracting depth information from a focal stack is disclosed. The method may include processing the focal stack through a focus convolutional neural network (Focus-Net) to generate a plurality of feature maps, stacking the plurality of feature maps together, and fusing the plurality of feature maps by a plurality of first convolutional layers to obtain a depth image. The Focus-Net includes a plurality of branches, and each branch includes a downsampling convolutional layer having a different stride for downsampling the focal stack and a deconvolutional layer for upsampling the focal stack.

Method and apparatus for estimating depth of field information

A method for generating hyperspectral data-cubes based on a plurality of hyperspectral light field (H-LF) images is disclosed. Each H-LF image may have a different view and a different spectral band. The method may include calculating a magnitude histogram, a direction histogram, and an overlapping histogram of oriented gradient for a plurality of pixels; developing a spectral-invariant feature descriptor by combining the magnitude histogram, the direction histogram, and the overlapping histogram of oriented gradient; obtaining a correspondence cost of the H-LF images based on the spectral-invariable feature descriptor; performing H-LF stereo matching on the H-LF images to obtain a disparity map of a reference view; and generating hyperspectral data-cubes by using the disparity map of the reference view. A bin in the overlapping histogram of oriented gradient may comprise overlapping ranges of directions.

Method and system for hyperspectral light field imaging

A method for generating high resolution multi-spectral light fields is disclosed. The method may include capturing a multi-perspective spectral image which includes a plurality of sub-view images; aligning and warping the sub-view images to obtain low resolution multi-spectral light fields; obtaining a high resolution dictionary and a low resolution dictionary; obtaining a sparse representation based on the low resolution multi-spectral light fields and the low resolution dictionary; and generating high resolution multi-spectral light fields with the sparse representation and the high resolution directory. Each sub-view image is captured with a different perspective and a different spectral range. The multi-perspective spectral image is obtain with one exposure.

Method and system for snapshot multi-spectral light field imaging

A method of calibrating a camera array comprising a plurality of cameras configured to capture a plurality of images to generate a panorama, wherein the relative positions among the plurality of cameras are constant, the method comprising: moving the camera array from a first position to a second position; measuring a homogeneous transformation matrix of a reference point on the camera array between the first position and the second position; capturing images at the first position and the second position by a first camera and a second camera on the camera array; and determining a homogenous transformation matrix between the first camera and the second camera based on the images captured by the first camera and the second camera at the first position and the second position. The method further comprises identifying a feature in the images taken by the first camera at the first position and the second position, and estimating a rotation of the first camera from the first position to the second position based on the feature.

Calibration method and apparatus for panoramic stereo video system

A method of generating a stereoscopic panorama is provided, the method comprising: processing a first right image, a second right image, a first left image, and a second left image to derive a right homography between the first right image and the second right image and a left homography between the first left image and the second left image; stitching the first right image with the second right image to generate a right panorama, the first left image with the second left image to generate a left panorama wherein the right homography is consistent with the left homography; and generating a stereo panorama using the right panorama and the left panorama.

Growing community of inventors

Shanghai, China

Jingyi Yu