Yuyan Li

Yuyan Li

Seattle, Washington, United States
241 followers 223 connections

Join to view profile

Activity

Join now to see all activity

Experience

  • Apple Graphic

    Apple

    Seattle, Washington, United States

  • -

    Sunnyvale, California, United States

  • -

    Sunnyvale, California, United States

  • -

    Sunnyvale, California, United States

Education

Publications

Patents

  • Multi-View consistency regularization for semantic interpretation of Equirectangular panoramas

    Filed US patent app. 17/545,673

    We invented a dense regression framework that estimates 360-degree (omni-directional) depth or
    segmentation maps from an Equi-Rectangular Projection (ERP) image. Our method follows a general
    encoder-decoder pipeline, which involves both convolutional layers and global attention layers. Our
    contributions are three-folds. First, we invented a distortion-free convolutional module designed to handle
    the varying distortion in 360 image across different regions. Second, we developed the…

    We invented a dense regression framework that estimates 360-degree (omni-directional) depth or
    segmentation maps from an Equi-Rectangular Projection (ERP) image. Our method follows a general
    encoder-decoder pipeline, which involves both convolutional layers and global attention layers. Our
    contributions are three-folds. First, we invented a distortion-free convolutional module designed to handle
    the varying distortion in 360 image across different regions. Second, we developed the self-attention
    the module which uses distortion-free image embedding to compute the appearance attention and use

    spherical distance to compute the positional attention. Third, we are the first to use transformer and self-
    attention architecture to solve 360 dense regression.

  • Method for omnidirectional dense regression for machine perception tasks via distortion-free CNN and spherical self-attention

    Filed US patent app. 16/836,290

    This invention introduces a novel regularization term to improve the performance of a deep neural network for semantic interpretation of equal-rectangular panorama images. Our approach utilizes the consistencies between different views of panorama images to reduce the needs of large amount of labelled ground truth data during training. Our innovation can be applied to various business areas such as building construction & maintenance, augmented & virtual reality businesses to reduce the costs.

Projects

  • Omnidirectional RGB-D Image Representation and Scene Understanding

    -

    Explored 360 omnidirectional geometry and representation. Developed deep learning-based solutions to address the problems of indoor layout estimation, monocular depth estimation, and stereo matching under 360 image domain. The proposed framework for monocular depth estimation outperformed the current state-of-the-arts by a margin.

    Adopted and customized vision transformer on 360 image representation, achieved top performance on scene understanding tasks such as depth prediction and…

    Explored 360 omnidirectional geometry and representation. Developed deep learning-based solutions to address the problems of indoor layout estimation, monocular depth estimation, and stereo matching under 360 image domain. The proposed framework for monocular depth estimation outperformed the current state-of-the-arts by a margin.

    Adopted and customized vision transformer on 360 image representation, achieved top performance on scene understanding tasks such as depth prediction and semantic segmentation.

  • 3D Point Cloud Feature Learning, Reconstruction, and Semantic Segmentation

    -

    Performed point cloud feature learning and representation on both large-scale indoor data and outdoor Lidar benchmarks. Designed and optimized deep learning and computer vision algorithms with a focus on efficient and effective performance for point cloud semantic segmentation and reconstruction tasks.

    Developed novel algorithms and designed CNN architectures based on both voxel and point convolution for the task of point cloud semantic segmentation. Achieved top-ranking performances in…

    Performed point cloud feature learning and representation on both large-scale indoor data and outdoor Lidar benchmarks. Designed and optimized deep learning and computer vision algorithms with a focus on efficient and effective performance for point cloud semantic segmentation and reconstruction tasks.

    Developed novel algorithms and designed CNN architectures based on both voxel and point convolution for the task of point cloud semantic segmentation. Achieved top-ranking performances in several challenging datasets, such as S3DIS, ScanNet, and SemanticKitti, etc.

  • Biomedical Image Synthesis

    -

    Used CNN, GAN architectures to synthesize unseen, high-quality, CT images for recovering full tomographic information from CT scans. The well-designed architecture performs optical flow estimation and images interpolation/extrapolation and receives state-of-the-art accuracy.

  • Multi-view RGB-D Image Registration and 3D Model Reconstruction

    -

    Implemented structure from motion, which is based on traditional SIFT feature matching, and bundle adjustment techniques to find correspondences between indoor multi-view image captured by Kinect v2 and reconstruct 3D scene represented as point cloud.

  • 3D Textured Mesh Model Reconstruction

    -

    Developed algorithms to convert 3D point clouds of buildings into simplified, high-quality mesh models with real-world image textures. Generated consistent building model texture by performing image stitching

Languages

  • Chinese

    Native or bilingual proficiency

  • English

    Full professional proficiency

More activity by Yuyan

View Yuyan’s full profile

  • See who you know in common
  • Get introduced
  • Contact Yuyan directly
Join to view full profile

Other similar profiles

Explore more posts

Explore collaborative articles

We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.

Explore More

Others named Yuyan Li in United States

10 others named Yuyan Li in United States are on LinkedIn

See others named Yuyan Li

Add new skills with these courses

See all courses